From Chatbots to Coworkers: How Worklayer Is Building Trust Infrastructure for AI Agents

Behind the code.

Behind [music] the code.

Behind [music] the code. That's where the future grows.

That's where the future grows. Welcome to Behind the Code brought to

Welcome to Behind the Code brought to you by Atomic Build. My name is Emit.

you by Atomic Build. My name is Emit. Today I have with me Milan from Work

Today I have with me Milan from Work Lair. Super excited to have you here

Lair. Super excited to have you here Milan. I'd love to hear an intro about

Milan. I'd love to hear an intro about yourself and your company. So I'll kick

yourself and your company. So I'll kick it over to you.

it over to you. Yeah. Hey, my name is Milan. I am

Yeah. Hey, my name is Milan. I am co-founder and CEO of Work Layer. and

co-founder and CEO of Work Layer. and worker we are building one password for

worker we are building one password for agents that's one way to describe what

agents that's one way to describe what we do and the second way is if a bit

we do and the second way is if a bit more technical crowd it's a inference

more technical crowd it's a inference environment for agents and I can go

environment for agents and I can go deeper into that in a second but my

deeper into that in a second but my background is I'm a machine learning

background is I'm a machine learning engineer I did I studied meatronics and

engineer I did I studied meatronics and I've been in this AI space since 2016 so

I've been in this AI space since 2016 so about a decade of experience here

about a decade of experience here wow that yeah I mean that is ahead of

wow that yeah I mean that is ahead of the curve when it comes to like

the curve when it comes to like understanding how AI systems work. And

understanding how AI systems work. And when you say inference for agents, if

when you say inference for agents, if you don't mind giving like what by the

you don't mind giving like what by the way, great toggle on your site, the

way, great toggle on your site, the technical non-technical site. I I love

technical non-technical site. I I love that. I think I mentioned to you before

that. I think I mentioned to you before when I reached out to you. If you were

when I reached out to you. If you were able to give two definitions, the

able to give two definitions, the non-technical first and the technical

non-technical first and the technical one after, that'd be great. I think

one after, that'd be great. I think it'll help people understand a bit

it'll help people understand a bit better what you're working on.

better what you're working on. Yeah. So one of the things that work

Yeah. So one of the things that work layer is building is this foundation

layer is building is this foundation layer so that agents can go and take

layer so that agents can go and take actions on their own behalf or on your

actions on their own behalf or on your behalf.

behalf. Mhm.

Mhm. And uh that's why we simplified the

And uh that's why we simplified the tagline to one password to agents. One

tagline to one password to agents. One password for agents.

password for agents. Right.

Right. The reason for that is that one password

The reason for that is that one password for agents basically means um the same

for agents basically means um the same way we are using one password or

way we are using one password or password manager. We have our

password manager. We have our credentials in there and then when we

credentials in there and then when we are on a website, it helps us inject

are on a website, it helps us inject that credential and use that login and

that credential and use that login and then we can do whatever we want in the

then we can do whatever we want in the app. As agents do more and more work for

app. As agents do more and more work for us in the background, they're not going

us in the background, they're not going to go and use websites the same way we

to go and use websites the same way we do, right? So there's different kinds of

do, right? So there's different kinds of credentials agents would need. There

credentials agents would need. There would be access token for your APIs. It

would be access token for your APIs. It could be just API key directly. It could

could be just API key directly. It could be some other form of session cookie,

be some other form of session cookie, etc. if it's using browser. So agents

etc. if it's using browser. So agents can use websites in different ways. That

can use websites in different ways. That means we want we had to create something

means we want we had to create something novel for the agents. And second thing

novel for the agents. And second thing is one password only stores for us

is one password only stores for us humans only stores the credential.

humans only stores the credential. That's the whole job. It's a vault

That's the whole job. It's a vault and you can share it with other people

and you can share it with other people etc. That's that's another feature. But

etc. That's that's another feature. But with agents you have to reimagine what

with agents you have to reimagine what that means because not only agents need

that means because not only agents need the credential they also can now take

the credential they also can now take actions. So you can design the

actions. So you can design the permission on each action that you can

permission on each action that you can say this agent can use this tool. I auto

say this agent can use this tool. I auto approve this one for this one agent has

approve this one for this one agent has to confirm with me before so I get

to confirm with me before so I get notification even though agent could be

notification even though agent could be working in the background that like some

working in the background that like some random time I would get this list of

random time I would get this list of notifications. So when we think of one

notifications. So when we think of one password for agents, we're almost

password for agents, we're almost underelling it, but it it's the easiest

underelling it, but it it's the easiest way for nontechnical folks to kind of

way for nontechnical folks to kind of wrap their head around that, okay, if

wrap their head around that, okay, if agents are going to be doing work in the

agents are going to be doing work in the background for us, this is the way that

background for us, this is the way that you can securely connect your agent to

you can securely connect your agent to other things, you can control everything

other things, you can control everything from one place. It's your wallet of apps

from one place. It's your wallet of apps that your agents can use. And not only

that your agents can use. And not only apps, it can then also put memories and

apps, it can then also put memories and skills and then you can take that across

skills and then you can take that across different agents. So if you were to

different agents. So if you were to explain this to a business leader who

explain this to a business leader who understands AI, they're using it,

understands AI, they're using it, they've asked their organization to

they've asked their organization to start using it, but they don't quite

start using it, but they don't quite understand the part about governing an

understand the part about governing an agent because they've never implemented

agent because they've never implemented an agent. How would you explain it to

an agent. How would you explain it to them to just for their first approach?

them to just for their first approach? That's a good Yeah, that's a good

That's a good Yeah, that's a good question. So if a business leader is

question. So if a business leader is trying to figure out how they can deploy

trying to figure out how they can deploy agents for their use case, they have a

agents for their use case, they have a lot of open questions. And my approach

lot of open questions. And my approach has always been first of all do they

has always been first of all do they know that the agent can do the things

know that the agent can do the things that they want it to do and there are

that they want it to do and there are business leaders who are very well

business leaders who are very well versed in that there are business

versed in that there are business leaders who don't know and need

leaders who don't know and need education in that. So my first point

education in that. So my first point would be okay if you don't even know

would be okay if you don't even know what tool calls are you know first we

what tool calls are you know first we need to show you that this agent is

need to show you that this agent is capable of doing this once company and

capable of doing this once company and business leaders are at that stage where

business leaders are at that stage where they understand that agents can

they understand that agents can basically take any actions and there's

basically take any actions and there's different levels of intelligence and you

different levels of intelligence and you want the model switching and all that

want the model switching and all that they come with a list of requirements

they come with a list of requirements they want to be able to not tie to any

they want to be able to not tie to any vendor right so we offer that they want

vendor right so we offer that they want access control they want to control

access control they want to control which employee can connect to what apps

which employee can connect to what apps and this is one of the big ones that um

and this is one of the big ones that um a lot of companies are still just trying

a lot of companies are still just trying to figure this out. And this is where we

to figure this out. And this is where we kind of win the conversation with them

kind of win the conversation with them because we can easily come and say for

because we can easily come and say for your you get this list of employees for

your you get this list of employees for each employee your admin accounts can

each employee your admin accounts can control which connectors are available

control which connectors are available in those connectors which permissions

in those connectors which permissions you want. So if you want your sales team

you want. So if you want your sales team to never be able to send out emails or

to never be able to send out emails or any companywide, you don't want anyone

any companywide, you don't want anyone to send out emails via agents, you can

to send out emails via agents, you can set that permission and you can allow

set that permission and you can allow the agents to then draft the emails

the agents to then draft the emails instead. And when an agent drafts the

instead. And when an agent drafts the email, they sit in your inbox, they

email, they sit in your inbox, they never get sent. So now the

never get sent. So now the responsibility part from a business

responsibility part from a business leader's point is like I want people to

leader's point is like I want people to use agents to draft emails. I don't want

use agents to draft emails. I don't want them to use to send the emails out

them to use to send the emails out because that that could be

because that that could be miscommunication that can happen there.

miscommunication that can happen there. So now the responsibility layer for

So now the responsibility layer for business leader now falls on each

business leader now falls on each individual employee and this is what

individual employee and this is what business leaders get is when you have

business leaders get is when you have agents and people working every agent

agents and people working every agent run is responsible and backs stop by a

run is responsible and backs stop by a human and you can set that permission

human and you can set that permission layer using work layer that's basically

layer using work layer that's basically the the cell that we do.

the the cell that we do. Got it. So I guess like a natural

Got it. So I guess like a natural question for me would be do people

question for me would be do people interpret it as a division of decisions

interpret it as a division of decisions meaning like what do they think of it

meaning like what do they think of it like okay what is the agent going to

like okay what is the agent going to decide what is work layer going to

decide what is work layer going to handle and where does my team come in or

handle and where does my team come in or like where does a human come in is this

like where does a human come in is this a is this a natural response and then

a is this a natural response and then how do you then give back feedback when

how do you then give back feedback when someone comes to that kind of

someone comes to that kind of yeah this is not what people think about

yeah this is not what people think about the first thing that they that they come

the first thing that they that they come to us they're not like I want to have

to us they're not like I want to have decision and control because people are

decision and control because people are not even at the stage where they've

not even at the stage where they've given employees freedom to use agents

given employees freedom to use agents and if they have it's very ad hoc. So

and if they have it's very ad hoc. So there are very few companies that would

there are very few companies that would have this issue because even if you let

have this issue because even if you let people use clot today out of the

people use clot today out of the workforce maybe 2 or 3% of the people

workforce maybe 2 or 3% of the people are connecting Gmail to it and sending

are connecting Gmail to it and sending out emails via it. So the problem is not

out emails via it. So the problem is not really experienced and the people who do

really experienced and the people who do connect claude to their Gmail they are

connect claude to their Gmail they are kind of well versed in it but imagine

kind of well versed in it but imagine now you roll it out to 98% and people

now you roll it out to 98% and people have no education on how to use it

have no education on how to use it that's where the real real challenge is

that's where the real real challenge is people want there's one kind of customer

people want there's one kind of customer that wants everyone to start using

that wants everyone to start using claude more and there's the education

claude more and there's the education piece that's missing and a part of it is

piece that's missing and a part of it is just like hey let me show you the

just like hey let me show you the security part is something you're

security part is something you're concerned about we already saw that but

concerned about we already saw that but let me show you what happens

let me show you what happens When you give everyone in your company

When you give everyone in your company an agent that talks to their tools,

an agent that talks to their tools, right? You can see the data behind how

right? You can see the data behind how people are using these systems, what

people are using these systems, what kind of work people are doing, you get

kind of work people are doing, you get the all that data as well as like just

the all that data as well as like just added bonus because now you can see how

added bonus because now you can see how employees are using different tools

employees are using different tools through the queries that they're making

through the queries that they're making to cloud. So there's a lot of benefit.

to cloud. So there's a lot of benefit. But the decision-m part is once you

But the decision-m part is once you start building a workflow, then those

start building a workflow, then those decisions starts coming in. Once once

decisions starts coming in. Once once you say okay I want this to be done this

you say okay I want this to be done this much by people this much by agents or I

much by people this much by agents or I want agents to be doing these things.

want agents to be doing these things. People always overshoot what agents

People always overshoot what agents should do. They always want the magic

should do. They always want the magic agent to do everything. And then you

agent to do everything. And then you have to give them the reality that okay

have to give them the reality that okay would you really trust an agent to do

would you really trust an agent to do this for you? And then their questions

this for you? And then their questions would be no no their answers would be

would be no no their answers would be no. And then you would you would say

no. And then you would you would say okay so then who is going to be

okay so then who is going to be responsible? Because if agent is doing

responsible? Because if agent is doing this part the work has to be fallen on a

this part the work has to be fallen on a human so that they can review and they

human so that they can review and they can approve XYZ that the agent is

can approve XYZ that the agent is submitting as the outcome and that's

submitting as the outcome and that's where I think people start thinking

where I think people start thinking about wow this is more complex than I

about wow this is more complex than I thought a human agent system is not as

thought a human agent system is not as simple but basically what we have done

simple but basically what we have done with workflow uh work layer is this

with workflow uh work layer is this building block of permissions decisions

building block of permissions decisions which human is responsible for which

which human is responsible for which agent who reports to who all of that we

agent who reports to who all of that we have kind of built in. So you can go and

have kind of built in. So you can go and play around and uh and configure that.

play around and uh and configure that. So in your like when you're rolling out

So in your like when you're rolling out with a company or or someone's adopting

with a company or or someone's adopting the system like you mentioned,

the system like you mentioned, everyone's at least touched Claude or

everyone's at least touched Claude or ChatgBT and some at least chat interface

ChatgBT and some at least chat interface back and forth, but very few integrate

back and forth, but very few integrate tools into the like into the actual

tools into the like into the actual layer of the agent. So are those folks

layer of the agent. So are those folks do they just become your natural

do they just become your natural champions? Like do you lean on them to

champions? Like do you lean on them to then pull in other people? Is that how

then pull in other people? Is that how you get it going inside an organization?

you get it going inside an organization? Yeah, to to get going with an org, we

Yeah, to to get going with an org, we basically need a person who is

basically need a person who is experiencing some pain point. So I can I

experiencing some pain point. So I can I can talk about some of the pain points

can talk about some of the pain points that some of our champions have

that some of our champions have experienced.

experienced. Sure.

Sure. One pain point is that hey I am the only

One pain point is that hey I am the only person who ends up like setting these

person who ends up like setting these systems up and I don't get any work done

systems up and I don't get any work done and now my job has become this. And then

and now my job has become this. And then we say okay what would you want? What

we say okay what would you want? What what what would you want here? and they

what what would you want here? and they say okay we want I want to be able to

say okay we want I want to be able to create these agents then people can

create these agents then people can subscribe to them right so I create a

subscribe to them right so I create a sales agent in my company and now other

sales agent in my company and now other people can subscribe to that sales agent

people can subscribe to that sales agent and once that is done I don't want to be

and once that is done I don't want to be going going and setting up in each

going going and setting up in each individual's computer right so here

individual's computer right so here there are a lot of platforms including

there are a lot of platforms including ours that offers that um then there are

ours that offers that um then there are different things is I am now I have now

different things is I am now I have now become the security admin for this

become the security admin for this because I understand this the most and

because I understand this the most and people want to use this and I have been

people want to use this and I have been setting up these permissions but now we

setting up these permissions but now we want to bring chat GPT and now I have to

want to bring chat GPT and now I have to manage cloud and chat GPT both. Is there

manage cloud and chat GPT both. Is there an easy way for me to just like you know

an easy way for me to just like you know manage all the different agents that

manage all the different agents that connect to one place. So then instead of

connect to one place. So then instead of going to different accounts and setting

going to different accounts and setting up your connectors and setting

up your connectors and setting permissions and you have to manage that

permissions and you have to manage that manually you basically get one one item

manually you basically get one one item one wallet and you can just say okay

one wallet and you can just say okay tagged claw everyone will use these

tagged claw everyone will use these tools these employees will be able to

tools these employees will be able to use these ones and this is the

use these ones and this is the permission across all of the agents. So

permission across all of the agents. So that's the second pain point where

that's the second pain point where when you have become that security admin

when you have become that security admin you need to be able to control that

you need to be able to control that across all the platforms and we kind of

across all the platforms and we kind of give you that easy way. to that software

give you that easy way. to that software layer and the final is also cloud agents

layer and the final is also cloud agents is what we're recently solving. So a lot

is what we're recently solving. So a lot of the teams that are advanced where a

of the teams that are advanced where a lot of consulting companies actually

lot of consulting companies actually have been recently using cloud a lot

have been recently using cloud a lot with skills because a lot of the

with skills because a lot of the consulting uh workflows especially when

consulting uh workflows especially when it comes to delivery can be done

it comes to delivery can be done extremely well by cloud co-work. So what

extremely well by cloud co-work. So what they want right now is they want to be

they want right now is they want to be able to automate a bunch of these

able to automate a bunch of these workflows that work really well. they

workflows that work really well. they have figured out we have these skills

have figured out we have these skills that work really well. the outputs and

that work really well. the outputs and inputs are most of the time correct and

inputs are most of the time correct and we have this person that will check at

we have this person that will check at the end and then you know we will submit

the end and then you know we will submit that draft or even after a meeting this

that draft or even after a meeting this is the session that claude is going to

is the session that claude is going to do with the person or with the BD uh rep

do with the person or with the BD uh rep who did the meeting and then based on

who did the meeting and then based on the notes it's going to go and then take

the notes it's going to go and then take the next actions right so those are

the next actions right so those are those kind of workflows they are

those kind of workflows they are something that people want to automate

something that people want to automate another example of automation that

another example of automation that people want to do is okay I I do

people want to do is okay I I do meetings and I I promise people that I

meetings and I I promise people that I will follow up with like X Y and Z item,

will follow up with like X Y and Z item, right? And right now they have a skill

right? And right now they have a skill that they run manually that says okay go

that they run manually that says okay go through my last four days and tell me

through my last four days and tell me what have I promised to people. Give me

what have I promised to people. Give me a list by person. So each person and

a list by person. So each person and then things that I need to do for that

then things that I need to do for that person and once you have that list it's

person and once you have that list it's very easy for you to go out in the real

very easy for you to go out in the real world and like do and take those actions

world and like do and take those actions or make introductions etc. So here

or make introductions etc. So here people run the scale randomly sometimes

people run the scale randomly sometimes they forget sometimes they don't. And

they forget sometimes they don't. And now this champion is saying can I just

now this champion is saying can I just like some people are old they are in

like some people are old they are in their 40s like late 40s 50s early 50s

their 40s like late 40s 50s early 50s and they've never like they were never

and they've never like they were never techsavvy but they have like great

techsavvy but they have like great consulting experience uh and that's

consulting experience uh and that's [clears throat] why they are in this in

[clears throat] why they are in this in this field and and we want them to be

this field and and we want them to be able to be up to date right and there's

able to be up to date right and there's nothing to do with agents it's just a

nothing to do with agents it's just a lot of people are techsavvy and then

lot of people are techsavvy and then they they do it that way a lot of people

they they do it that way a lot of people don't but here he wants everyone to be

don't but here he wants everyone to be able to use this and I I think a better

able to use this and I I think a better UX there is can you just get an email

UX there is can you just get an email every day of the promises you have made

every day of the promises you have made and promises you have kept and then as

and promises you have kept and then as you fulfill those promises you can check

you fulfill those promises you can check those off and that that workflow doing

those off and that that workflow doing that with the agent chat is not that

that with the agent chat is not that great. So having a cloud agent just kind

great. So having a cloud agent just kind of do that you text it back that like

of do that you text it back that like hey I've completed these things and and

hey I've completed these things and and you get that in the text message that is

you get that in the text message that is the type of work that like now people

the type of work that like now people are looking for is is that endstate

are looking for is is that endstate automation on what is the good UX for a

automation on what is the good UX for a person to provide feedback what is a

person to provide feedback what is a good interface for the agent to keep

good interface for the agent to keep doing work and if agent is doing work in

doing work and if agent is doing work in the background and if it's a dumb

the background and if it's a dumb automation as in it just sends me

automation as in it just sends me notification every morning and if I

notification every morning and if I don't look at that it's a useless not

don't look at that it's a useless not it's a useless automation how do to make

it's a useless automation how do to make it smart, right? Like how do you make

it smart, right? Like how do you make the agent? And this is where the open

the agent? And this is where the open claw style heartbeat comes into place.

claw style heartbeat comes into place. If you have not responded for the last 3

If you have not responded for the last 3 days, it can just nag you a little bit

days, it can just nag you a little bit and say, "Okay, let's at least get like

and say, "Okay, let's at least get like two follow-ups done so that I can mark

two follow-ups done so that I can mark this done and and it has intelligence

this done and and it has intelligence layer there where where its goal is to

layer there where where its goal is to help you clear out the follow-ups and

help you clear out the follow-ups and now you are work now now you're creating

now you are work now now you're creating an agent that is more uh designed for a

an agent that is more uh designed for a human than more designed for an

human than more designed for an automation. So those are the kind of

automation. So those are the kind of problems that are for the frontier

problems that are for the frontier people who are you know automating a lot

people who are you know automating a lot with their skills but everything is a

with their skills but everything is a routine and and a dumb automation and

routine and and a dumb automation and they wanted to be a smart agent.

they wanted to be a smart agent. So I mean that's I think I think you

So I mean that's I think I think you just to unpack some of what you said

just to unpack some of what you said there was a lot there. So the AI fluency

there was a lot there. So the AI fluency changes from generations, right? Like

changes from generations, right? Like some people are coming into the

some people are coming into the workforce. It's you have to be AI fluent

workforce. It's you have to be AI fluent day one. Others have existed for 10, 20,

day one. Others have existed for 10, 20, 30 years. But you're mentioning that

30 years. But you're mentioning that there's a graduation from creating a

there's a graduation from creating a skill that comes back to you and you

skill that comes back to you and you actually turn to an editor to now almost

actually turn to an editor to now almost like you accept the agent can be an

like you accept the agent can be an author. What what have like what have

author. What what have like what have your users had that moment or shared

your users had that moment or shared with you like this is what made me think

with you like this is what made me think this is possible? Is it just reps? Like

this is possible? Is it just reps? Like they're using it often enough to see

they're using it often enough to see this can be autonomous. It doesn't need

this can be autonomous. It doesn't need my like editing all the time. Is that

my like editing all the time. Is that one of the key things? I'm curious cuz

one of the key things? I'm curious cuz for me that's what it was. It was just

for me that's what it was. It was just reps that unlocked it for me.

reps that unlocked it for me. Well, so reps is something that

Well, so reps is something that definitely helps and that's how people

definitely helps and that's how people have been doing it. We've been trying to

have been doing it. We've been trying to figure out a way without reps like how

figure out a way without reps like how do you give how do you diffuse knowledge

do you give how do you diffuse knowledge within a company without reps and

within a company without reps and I think a lot of a lot of people have

I think a lot of a lot of people have talked about this which is

talked about this which is put a agent in Slack and let people use

put a agent in Slack and let people use it and let other people see how other

it and let other people see how other people are using it and I think that's

people are using it and I think that's the biggest way to do knowledge

the biggest way to do knowledge diffusion of what agents can do and

diffusion of what agents can do and cannot do. So the champions who are

cannot do. So the champions who are wellversed in what kind of skills they

wellversed in what kind of skills they are the only ones who build skills other

are the only ones who build skills other people are not building skills still and

people are not building skills still and what this person does is they put this

what this person does is they put this slack agent which is very easy to create

slack agent which is very easy to create on our platform. You

on our platform. You one click deploy to Slack. So now you

one click deploy to Slack. So now you have this custom agent. Let's say it's

have this custom agent. Let's say it's doing uh CRM updates and this person

doing uh CRM updates and this person will go and say okay I just did this

will go and say okay I just did this meeting update these notes based on this

meeting update these notes based on this meeting from granola right and then it

meeting from granola right and then it would go and look into it and the cool

would go and look into it and the cool thing is because we are the one password

thing is because we are the one password layer when the person when I talk to

layer when the person when I talk to that agent it uses my granola or my

that agent it uses my granola or my fathom or my Google meets when you use

fathom or my Google meets when you use the same agent in the same slack it will

the same agent in the same slack it will use your apps right

use your apps right and basically you get this control of

and basically you get this control of like okay it's the same agent but it has

like okay it's the same agent but it has aware of like which person is using is

aware of like which person is using is and there's also agent credentials. So

and there's also agent credentials. So sometimes you don't want agent to use

sometimes you don't want agent to use user's credential but you have you have

user's credential but you have you have want to give agent their own credential.

want to give agent their own credential. It could be their own email ID, it could

It could be their own email ID, it could be their own Google drive. So that's

be their own Google drive. So that's shared and we also offer that. So those

shared and we also offer that. So those are the kind of things that I think are

are the kind of things that I think are the biggest leap when it comes to people

the biggest leap when it comes to people just kind of expanding their horizons on

just kind of expanding their horizons on what agents can do.

what agents can do. Yeah. In a company that is the best way

Yeah. In a company that is the best way to diffuse knowledge

to diffuse knowledge to to like have it on display like like

to to like have it on display like like display multiplayer so one person can

display multiplayer so one person can message second person can chime in and

message second person can chime in and and imagine like what happens at a

and imagine like what happens at a company so this consulting company that

company so this consulting company that we're working with a lot of people are

we're working with a lot of people are you know experts that they have worked

you know experts that they have worked for over 30 plus years of experience

for over 30 plus years of experience they're very good and very sharp with

they're very good and very sharp with like the knowledge that they have

like the knowledge that they have accumulated they don't know anything

accumulated they don't know anything about what agents can do and then there

about what agents can do and then there are young people who are in charge of

are young people who are in charge of doing a lot of putting up the decks and

doing a lot of putting up the decks and all that together and they have built

all that together and they have built the skills to say okay go and build this

the skills to say okay go and build this deck out based on like our workshop

deck out based on like our workshop session very quickly and this person

session very quickly and this person didn't have to build any skill anything

didn't have to build any skill anything it just had the slack agent available to

it just had the slack agent available to just tag it and say okay now can you

just tag it and say okay now can you build a SL deck for me and this works

build a SL deck for me and this works extremely well because

extremely well because I don't think everyone needs to learn

I don't think everyone needs to learn how to use agents anyways

how to use agents anyways it just needs to be available and you

it just needs to be available and you can start using it it's just it just

can start using it it's just it just needs to be there when you need it and

needs to be there when you need it and Slack is one great way to do it.

Slack is one great way to do it. iMessage is another one, but iMessage is

iMessage is another one, but iMessage is human to human and it's more like

human to human and it's more like personal assistant. But the surface

personal assistant. But the surface really matters. If you do it Discord,

really matters. If you do it Discord, it's more community. If you do it Slack,

it's more community. If you do it Slack, it's more internal use. You can also

it's more internal use. You can also have community and customer support

have community and customer support agents on Slack with customer channels.

agents on Slack with customer channels. So there's all these different I think

So there's all these different I think just the the layer of like where people

just the the layer of like where people use it changes what the agent can

use it changes what the agent can actually do. It's it's very cool.

actually do. It's it's very cool. Yeah. So, so would you say like

Yeah. So, so would you say like something has to change inside

something has to change inside organizations for this to become like an

organizations for this to become like an isolated productivity game to more

isolated productivity game to more companywide and for you like the single

companywide and for you like the single biggest unlock is making it accessible

biggest unlock is making it accessible like truly accessible in a very simple

like truly accessible in a very simple way like Slack is probably one of the

way like Slack is probably one of the most simple way an internal team member

most simple way an internal team member can access it and then from there I

can access it and then from there I guess it's just leaning into it and

guess it's just leaning into it and trying like would you say just like you

trying like would you say just like you have to make the leap there is nothing

have to make the leap there is nothing in there's no in between aside from

in there's no in between aside from making the leap

making the leap I don't think so like the Slack thing

I don't think so like the Slack thing has been crazy in terms of adoption.

has been crazy in terms of adoption. Once a Slack agent exists, the number of

Once a Slack agent exists, the number of messages sent to an agent like chart

messages sent to an agent like chart spikes very very fast. So we have we

spikes very very fast. So we have we think that this is our our wedge. We are

think that this is our our wedge. We are going to eventually do Microsoft Teams

going to eventually do Microsoft Teams once the Slack thing has been working

once the Slack thing has been working really well and we have exhausted it.

really well and we have exhausted it. But right now just so many companies

But right now just so many companies already use Slack.

already use Slack. There's a big market right just right

There's a big market right just right there and I don't think we have

there and I don't think we have exhausted all the different techniques

exhausted all the different techniques just with Slack yet. So we are still

just with Slack yet. So we are still exploring like okay how do we even like

exploring like okay how do we even like spread this adoption like wildfire

spread this adoption like wildfire on top of Slack but right now it just

on top of Slack but right now it just like Slack is bringing so much usage

like Slack is bringing so much usage that yeah it's it's interesting.

that yeah it's it's interesting. Do do you see a nearby future where the

Do do you see a nearby future where the Slack agent is one of the champions

Slack agent is one of the champions acting on behalf of of work layer to get

acting on behalf of of work layer to get people involved and starting using it

people involved and starting using it inside their company? Yeah, we are

inside their company? Yeah, we are exploring this uh agent that is just

exploring this uh agent that is just going to be their company brand that you

going to be their company brand that you can also install and it becomes this

can also install and it becomes this entry point. So you don't need to create

entry point. So you don't need to create your custom Slackbot or Slack agent.

your custom Slackbot or Slack agent. Just install this from the market and it

Just install this from the market and it has all the information on how to set up

has all the information on how to set up our account or or companies account. You

our account or or companies account. You can create custom agents and then you

can create custom agents and then you can also use it as a personal assistant.

can also use it as a personal assistant. So everyone but this is very similar to

So everyone but this is very similar to what Claude does. So in product sense it

what Claude does. So in product sense it might just feel like okay it's a copy

might just feel like okay it's a copy product but really what we're trying to

product but really what we're trying to build in is okay how do you really use

build in is okay how do you really use work layer and and and I I wouldn't call

work layer and and and I I wouldn't call it more of a internal AI champion even

it more of a internal AI champion even though that's a really good idea we

though that's a really good idea we should really look into it that instead

should really look into it that instead of a person being that champion it can

of a person being that champion it can be an agent it is this forward deployed

be an agent it is this forward deployed engineer the way we are describing it is

engineer the way we are describing it is you can bring your problems to it and it

you can bring your problems to it and it will give you ideas on how to solve it

will give you ideas on how to solve it right now the AI champions do a little

right now the AI champions do a little bit more than that they're not forward

bit more than that they're not forward deployed engineers they are basically

deployed engineers they are basically going and building this out and then

going and building this out and then helping you like teach it and and doing

helping you like teach it and and doing the full loop versus this agent is just

the full loop versus this agent is just kind of giving you information. It would

kind of giving you information. It would be nice if it could go back and build

be nice if it could go back and build everything and then do what the forward

everything and then do what the forward deploy engineers actually do or or these

deploy engineers actually do or or these AI champions actually do. So

AI champions actually do. So yeah. Yeah. No, that's a great point. I

yeah. Yeah. No, that's a great point. I think so. Human in the loop has been

think so. Human in the loop has been used predates like widespread usage of

used predates like widespread usage of AI, but it seems like we're going into

AI, but it seems like we're going into this era where it's like a bridge until

this era where it's like a bridge until agents improve to a point where they're

agents improve to a point where they're self-governing and they can like perform

self-governing and they can like perform at a very high level. But do you see it

at a very high level. But do you see it as a permanent feature or of like

as a permanent feature or of like responsible agent systems in companies

responsible agent systems in companies where there will always be a human actor

where there will always be a human actor in the loop?

in the loop? This this is a very philosophical on how

This this is a very philosophical on how you believe the world works. Yeah,

you believe the world works. Yeah, for me the world works because someone

for me the world works because someone is paying for something, right? right?

is paying for something, right? right? Like the economy basically

Like the economy basically it works because a person is paying for

it works because a person is paying for something convenience services goods

something convenience services goods commodity and uh at the end of the day

commodity and uh at the end of the day wallet is with a person right so even if

wallet is with a person right so even if like you're building a product that is

like you're building a product that is used by other AI agents at the end of

used by other AI agents at the end of the day behind that person is a wallet

the day behind that person is a wallet behind that agent is a person with a

behind that agent is a person with a wallet that's paying for you know so

wallet that's paying for you know so let's say you are you have a product

let's say you are you have a product that you know agents can use and pay you

that you know agents can use and pay you to use your product like it could be a

to use your product like it could be a search tool still at the end of the day

search tool still at the end of the day there's as a person is is what my point

there's as a person is is what my point is. So

is. So the way I see your question is the human

the way I see your question is the human in the loop part let's say Amazon or any

in the loop part let's say Amazon or any company is more and more getting

company is more and more getting automated or even a small company with

automated or even a small company with with like 10 employees is automating.

with like 10 employees is automating. What I see is that yes that company will

What I see is that yes that company will be able to do more. I don't see one

be able to do more. I don't see one person company really. I can see 10

person company really. I can see 10 people company running because you have

people company running because you have to think psychologically one person

to think psychologically one person company is very lonely. No, no one

company is very lonely. No, no one cares.

cares. It's very lonely. It's it's if something

It's very lonely. It's it's if something breaks, let's say as a oneperson company

breaks, let's say as a oneperson company now that the bar has raised been raised

now that the bar has raised been raised that what a company can do because

that what a company can do because everyone's using agents. If a one person

everyone's using agents. If a one person company tries to survive in that, you

company tries to survive in that, you have all these automated systems working

have all these automated systems working for you. If one fails, okay, you can pay

for you. If one fails, okay, you can pay attention to that and fix it. If two

attention to that and fix it. If two fails, three fails, you have to go fix

fails, three fails, you have to go fix three at a row. And now think of like

three at a row. And now think of like how big your operation is. If 100 fails

how big your operation is. If 100 fails at the same time, you are the sole

at the same time, you are the sole responsible to fix those 100 things. And

responsible to fix those 100 things. And it's very difficult. your s your servers

it's very difficult. your s your servers could go down, systems could go down. So

could go down, systems could go down. So how people will create companies and how

how people will create companies and how people will put agents in the work is

people will put agents in the work is yes one person will be responsible to do

yes one person will be responsible to do a lot of things but you would design it

a lot of things but you would design it almost in a way where if cascading

almost in a way where if cascading failures do happen you still have enough

failures do happen you still have enough people responsible in the layer that can

people responsible in the layer that can take those things out right so you would

take those things out right so you would have the agent layer which will be just

have the agent layer which will be just agents doing work and then you would

agents doing work and then you would have human layer and one simple way is

have human layer and one simple way is to think one person responsible for this

to think one person responsible for this 100 agents but that's not the way you

100 agents but that's not the way you might just keep your agent layer with

might just keep your agent layer with like a million agents and you would have

like a million agents and you would have a 100 people hired who are basically on

a 100 people hired who are basically on call and whenever something breaks they

call and whenever something breaks they go and fix that right and generally

go and fix that right and generally something is breaking because there's

something is breaking because there's million agents going on so that's one

million agents going on so that's one way human in the loop does come into

way human in the loop does come into play even if there's full automation you

play even if there's full automation you would have that the human in the loop

would have that the human in the loop people do today is like letting agent

people do today is like letting agent use a tool not use a tool or uh

use a tool not use a tool or uh basically agent asks a question and then

basically agent asks a question and then you you respond and all that stuff I can

you you respond and all that stuff I can see that can got automated away because

see that can got automated away because at the end of the day it's just data

at the end of the day it's just data right now agents are collecting by

right now agents are collecting by asking you questions. Once that is

asking you questions. Once that is formed into a skill, it can be it can be

formed into a skill, it can be it can be created in a way where okay,

created in a way where okay, intelligence is good enough, tools are

intelligence is good enough, tools are stable enough, harness is durable

stable enough, harness is durable enough, this is all going to run behind

enough, this is all going to run behind the scenes. But when failures do happen,

the scenes. But when failures do happen, that's like the real human in the loop

that's like the real human in the loop that I think and I don't think there's a

that I think and I don't think there's a there's a world where people go away. We

there's a world where people go away. We just will find like the higher like our

just will find like the higher like our dreams and ambitions are bigger than

dreams and ambitions are bigger than AI's dreams and ambitions today. I don't

AI's dreams and ambitions today. I don't know that in 10 years if AI is are able

know that in 10 years if AI is are able to set their own goals and work on their

to set their own goals and work on their own thing and they have their own

own thing and they have their own economy. They have wallets now. That's a

economy. They have wallets now. That's a different world. That's where the scary

different world. That's where the scary part comes in. But but in this world

part comes in. But but in this world right now, at least for the la next 5 to

right now, at least for the la next 5 to 10 years, I just don't see agents

10 years, I just don't see agents getting their own like identity, goals,

getting their own like identity, goals, money, and navigation of the real world.

money, and navigation of the real world. Like even if they do that, they're not

Like even if they do that, they're not in a robot body today. Robots are not

in a robot body today. Robots are not that good enough where they can be their

that good enough where they can be their own entity. at the end of the day they

own entity. at the end of the day they can only do stuff on the computer and

can only do stuff on the computer and they'll still come to you and say hey I

they'll still come to you and say hey I need this work done and they might pay

need this work done and they might pay you for that to go and take action in

you for that to go and take action in real world and that's still human in the

real world and that's still human in the loop so

loop so there's like very interesting ways to

there's like very interesting ways to think about it but yeah I I I am a very

think about it but yeah I I I am a very much of a person who thinks that we as

much of a person who thinks that we as people are always good enough to

people are always good enough to navigate the world the worry of like us

navigate the world the worry of like us not having any job to do is true but

not having any job to do is true but it's only transitionary it's it's right

it's only transitionary it's it's right now it's true we'll find something else

now it's true we'll find something else to do very soon.

to do very soon. So I guess like your just to summarize

So I guess like your just to summarize and correct me if I'm wrong, your take

and correct me if I'm wrong, your take is like as long as there's an economic

is like as long as there's an economic imperative for a human to even to act

imperative for a human to even to act with machines, they'll always be

with machines, they'll always be something to do in the broader economy

something to do in the broader economy which is like I think

which is like I think yeah okay no that makes sense and one

yeah okay no that makes sense and one thing you were saying in one of the just

thing you were saying in one of the just now was that if you were a one person

now was that if you were a one person company and you have to govern all these

company and you have to govern all these actions things break and they cascade. I

actions things break and they cascade. I remember I can't remember where it

remember I can't remember where it posted or where you presented about

posted or where you presented about self-healing code like self-healing

self-healing code like self-healing engineering workflows. I'd love for you

engineering workflows. I'd love for you just to explain that quickly and even if

just to explain that quickly and even if you can share an example of self-healing

you can share an example of self-healing codebase like in your own work. I think

codebase like in your own work. I think that'd be really good for people to

that'd be really good for people to hear.

hear. Yeah. So self-healing codebase is

Yeah. So self-healing codebase is basically agent running in a loop every

basically agent running in a loop every morning. It goes looks at production

morning. It goes looks at production data, goes looks at feedback users have

data, goes looks at feedback users have submitted on the app. uh it goes looks

submitted on the app. uh it goes looks at our live QA testing that we have done

at our live QA testing that we have done asynchronously throughout the the

asynchronously throughout the the previous day. It takes all that

previous day. It takes all that synthesizes it dumps it into a database.

synthesizes it dumps it into a database. Right now we're using notion for that.

Right now we're using notion for that. It's a very simple process from post

It's a very simple process from post talk to notion and there's the skill in

talk to notion and there's the skill in between that does the the the conversion

between that does the the the conversion right because post talk just has logs

right because post talk just has logs and then you want them converted to

and then you want them converted to issues. This is not that difficult. The

issues. This is not that difficult. The second part is also another agent

second part is also another agent running autonomously in a loop. every

running autonomously in a loop. every 400 p.m. it goes and checks the

400 p.m. it goes and checks the database, notion database. If anything

database, notion database. If anything is marked for the agent to do, it will

is marked for the agent to do, it will pick that out, spin up a PR, complete

pick that out, spin up a PR, complete the PR, test it, and then let it put it

the PR, test it, and then let it put it on a PR stack for us to review. So, it's

on a PR stack for us to review. So, it's these two things that I presented as

these two things that I presented as like this one entire system called

like this one entire system called self-healing codebase that if something

self-healing codebase that if something breaks in production, user experiences

breaks in production, user experiences an error, automatically converts to

an error, automatically converts to issue, that issue gets picked up and

issue, that issue gets picked up and that gets we had a human in the loop

that gets we had a human in the loop there. So the morning agent runs at 9:00

there. So the morning agent runs at 9:00 a.m. and then the evening one runs at 4

a.m. and then the evening one runs at 4 p.m. And if there are like there's one

p.m. And if there are like there's one person who goes in, there's a dashboard

person who goes in, there's a dashboard that we have built and you basically can

that we have built and you basically can press one, two or three as a button. So

press one, two or three as a button. So if you press one, you say assign to

if you press one, you say assign to agent. If you press two, assign to a

agent. If you press two, assign to a user. Three, this is not really an

user. Three, this is not really an issue. Let's discard it. And you can

issue. Let's discard it. And you can just press that one, two, three button,

just press that one, two, three button, and the issues will like keep popping

and the issues will like keep popping up. We'd build this UI, a special review

up. We'd build this UI, a special review UI for a person to go in and quickly

UI for a person to go in and quickly annotate which and they can also add a

annotate which and they can also add a comment if they want there with just

comment if they want there with just like pressing a button C. Right? So we

like pressing a button C. Right? So we we made it very frictionless for the

we made it very frictionless for the person to go in and just review these

person to go in and just review these issues that are being queued up by the

issues that are being queued up by the agent. So human is the bottleneck here

agent. So human is the bottleneck here because we didn't want to waste tokens

because we didn't want to waste tokens on issues that should not be solved,

on issues that should not be solved, right? And dduping the issues is the

right? And dduping the issues is the work of the first agent. Making PRs is

work of the first agent. Making PRs is the work of the second agent, but the

the work of the second agent, but the second agent only acts on items that are

second agent only acts on items that are queued for the agent. So even the first

queued for the agent. So even the first one like puts 100 items in, if those 100

one like puts 100 items in, if those 100 items are not reviewed by a human, the

items are not reviewed by a human, the second one doesn't pick up. Do we get to

second one doesn't pick up. Do we get to a point where the human is not needed

a point where the human is not needed when the agent is really working well?

when the agent is really working well? Like both of these agents are working

Like both of these agents are working well. Yes, we can remove that human in

well. Yes, we can remove that human in the loop review item. At the end of the

the loop review item. At the end of the day, you still review the PRs that

day, you still review the PRs that you're merging in because you care about

you're merging in because you care about the code. Now at some point you can say

the code. Now at some point you can say okay agents are so good at writing code

okay agents are so good at writing code you don't even need that sure that's

you don't even need that sure that's great now your bugs are completely

great now your bugs are completely fixing automatically and you're not

fixing automatically and you're not doing anything and the only thing maybe

doing anything and the only thing maybe you're doing is monitoring the system

you're doing is monitoring the system and checking if like which bugs it's

and checking if like which bugs it's picking which one it's fixing why it's

picking which one it's fixing why it's making this decisions etc but at the end

making this decisions etc but at the end of the day you're still there because

of the day you're still there because the software that I am building is for a

the software that I am building is for a user and if our users start complaining

user and if our users start complaining that this is not working or these

that this is not working or these features are you know bloating the

features are you know bloating the product that's still a feedback that

product that's still a feedback that comes as a meta channel. So there's all

comes as a meta channel. So there's all these different things. But yeah, that's

these different things. But yeah, that's basically what our self-healing codebase

basically what our self-healing codebase was. It was a demo of showing people

was. It was a demo of showing people that look, this is the starting and this

that look, this is the starting and this was a few months ago. I think this was

was a few months ago. I think this was during Toronto Tech Week in May, end of

during Toronto Tech Week in May, end of May.

May. Yeah,

Yeah, it's August.

it's August. That's right.

That's right. Oh yeah, early June. Now it's August. So

Oh yeah, early June. Now it's August. So people have been doing this more and

people have been doing this more and more now. So this might not feel that

more now. So this might not feel that crazy today but in May when we did this

crazy today but in May when we did this we really felt that well there's this

we really felt that well there's this software factory that people can create

software factory that people can create but you really need to have control like

but you really need to have control like some issues we experienced were PR's

some issues we experienced were PR's getting stacked and no one has time to

getting stacked and no one has time to review it and then we had to like break

review it and then we had to like break the loop and then other one was like

the loop and then other one was like issues are coming in but they weren't

issues are coming in but they weren't being dduped as in agent found this

being dduped as in agent found this issue today logged it next day same

issue today logged it next day same issue comes in logs it as a separate

issue comes in logs it as a separate issue and now you have two same issues

issue and now you have two same issues in the dashboard and the way we solved

in the dashboard and the way we solved it is like okay let's just improve the

it is like okay let's just improve the skill on the issue side let's improve

skill on the issue side let's improve the skill on this implementation side

the skill on this implementation side and one of the things that we added was

and one of the things that we added was a control loop so if there are more than

a control loop so if there are more than five PRs waiting the implementation

five PRs waiting the implementation agent would just say hey you are not

agent would just say hey you are not like picking up the work so I'm not

like picking up the work so I'm not going to do the work right so you can

going to do the work right so you can design these kind of controls in your

design these kind of controls in your loop so you don't waste tokens and these

loop so you don't waste tokens and these are the kind of things that I think

are the kind of things that I think about a lot is every task you want

about a lot is every task you want automatic outcomes but you also want to

automatic outcomes but you also want to like control when those outcomes keep

like control when those outcomes keep happening if humans are not reviewing it

happening if humans are not reviewing it what's the point of like doing that

what's the point of like doing that outcome because no like that that work

outcome because no like that that work is what I call slop productivity right

is what I call slop productivity right you're just like generating slop no

you're just like generating slop no one's checking and it's just getting

one's checking and it's just getting parked and

parked and uh it's useless

uh it's useless so part of token efficiency from your

so part of token efficiency from your perspective is to actually pull people

perspective is to actually pull people into the workflow consistently to ensure

into the workflow consistently to ensure they're engaging essentially

they're engaging essentially today yeah because uh you can create

today yeah because uh you can create closed loop systems right so in the

closed loop systems right so in the concept of control theory there is

concept of control theory there is openloop systems and closed loop systems

openloop systems and closed loop systems openloop systems are where a system

openloop systems are where a system happens and feedback is not received

happens and feedback is not received back in the in the next run. In closed

back in the in the next run. In closed loop systems, some work happens,

loop systems, some work happens, feedback comes in via the sensor and

feedback comes in via the sensor and then basically the uh the system gets

then basically the uh the system gets stabilized. So airplanes balancing

stabilized. So airplanes balancing themselves out. It's a PD controller.

themselves out. It's a PD controller. It's a control loop because every sensor

It's a control loop because every sensor information is coming back in and it's

information is coming back in and it's taking the action. So if it's if it

taking the action. So if it's if it moves this way too much, it can move

moves this way too much, it can move back and and then it can stabilize. Same

back and and then it can stabilize. Same way any stabilizer system like camera

way any stabilizer system like camera gyros, they work this way. So if you

gyros, they work this way. So if you take that same concept out, there are

take that same concept out, there are things where you can do closed loop

things where you can do closed loop systems with agents. If a really good

systems with agents. If a really good example is um are you familiar with

example is um are you familiar with front end and react as a framework?

front end and react as a framework? Okay, so people use React to create

Okay, so people use React to create frontends. There's this really cool test

frontends. There's this really cool test library called React Doctor uh made by

library called React Doctor uh made by this guy on on Twitter. He's uh he makes

this guy on on Twitter. He's uh he makes a lot of like interesting front end

a lot of like interesting front end tools. React Doctor gives your codebase

tools. React Doctor gives your codebase a score of how good your React system

a score of how good your React system is, how efficient it is, what what like

is, how efficient it is, what what like good and bad and it's a really good

good and bad and it's a really good library, right? Like you can run it

library, right? Like you can run it against your entire React app can give

against your entire React app can give you a score. Now you can think of

you a score. Now you can think of creating an agent a closed loop system

creating an agent a closed loop system where every morning it needs to go

where every morning it needs to go analyze and find any inefficiencies fix

analyze and find any inefficiencies fix that run react doctor and if react

that run react doctor and if react doctor increases the score and this

doctor increases the score and this issue is fixed and it's good go and

issue is fixed and it's good go and merge automatically and the whole goal

merge automatically and the whole goal of the system is to increase the react

of the system is to increase the react health score of your system and if you

health score of your system and if you let this agent run all it's doing is

let this agent run all it's doing is optimizing the front-end code you have

optimizing the front-end code you have given it instructions to not change any

given it instructions to not change any functionality you have put a adversarial

functionality you have put a adversarial review agent who doesn't implement it

review agent who doesn't implement it but just goes and checks the PR after

but just goes and checks the PR after implementation to check if this is

implementation to check if this is changing any functionality for the user

changing any functionality for the user if it doesn't then it can go merge it so

if it doesn't then it can go merge it so now this is a fully closed loop system

now this is a fully closed loop system and you can set this system just on the

and you can set this system just on the go and not have to worry about it at all

go and not have to worry about it at all right so if there are verifiable results

right so if there are verifiable results and and you can have a score and a

and and you can have a score and a metric that agent can monitor and keep

metric that agent can monitor and keep improving this is the perfect this is

improving this is the perfect this is what people do with slash goal or or

what people do with slash goal or or other type of things right like they're

other type of things right like they're running agents for uh days or weeks And

running agents for uh days or weeks And I have not seen any like good use other

I have not seen any like good use other than programming for this because

than programming for this because programming is the only one where you

programming is the only one where you can you can close the loop and verify

can you can close the loop and verify the rewards. But any optimization like

the rewards. But any optimization like your web app is taking like a few

your web app is taking like a few seconds to load throw agent at it and it

seconds to load throw agent at it and it can figure out and then at the end of

can figure out and then at the end of the day you can just look at the code

the day you can just look at the code and say this is up to my code quality or

and say this is up to my code quality or not. Uh but that's that's like the only

not. Uh but that's that's like the only thing you care about. If you don't even

thing you care about. If you don't even care about code quality and you just

care about code quality and you just assume like in this agent world that

assume like in this agent world that code is just like assembly, I don't

code is just like assembly, I don't really want to look at it. As long as

really want to look at it. As long as the system's working and no one's

the system's working and no one's complaining, I'm just going to throw it.

complaining, I'm just going to throw it. It's it's great. And I think we we'll

It's it's great. And I think we we'll get to that world, right? We'll not be

get to that world, right? We'll not be looking at our code. A lot of people

looking at our code. A lot of people don't look at their code today.

don't look at their code today. Yeah. Yeah, that's a good point. I think

Yeah. Yeah, that's a good point. I think I've seen examples of folks using Devon,

I've seen examples of folks using Devon, the engineering AI agent, just like

the engineering AI agent, just like autonomously. It's just pushing and

autonomously. It's just pushing and merging without much oversight. But like

merging without much oversight. But like to your point there I think there still

to your point there I think there still has to be like a rubric or a benchmark

has to be like a rubric or a benchmark to which things are measured. So maybe

to which things are measured. So maybe there's a future where

there's a future where those are also agent driven or maybe

those are also agent driven or maybe that's now maybe that's now maybe people

that's now maybe that's now maybe people are measuring against agent rubrics or

are measuring against agent rubrics or measurement systems.

measurement systems. Yeah I mean this becomes the job right

Yeah I mean this becomes the job right if we are trying to say what what will

if we are trying to say what what will people do is people will design how do

people do is people will design how do you give feedback to the agents. So

you give feedback to the agents. So mhm

mhm and every time you can build like you

and every time you can build like you can even think about it this way

can even think about it this way openloop systems as in agents doing work

openloop systems as in agents doing work in a openloop system will be more costly

in a openloop system will be more costly because human is in the review step

because human is in the review step somewhere at least once a closed loop

somewhere at least once a closed loop system like the react doctor thing they

system like the react doctor thing they will be commoditized to just the cost of

will be commoditized to just the cost of goods because

goods because a closed loop system at the end of the

a closed loop system at the end of the day doesn't need a person which means

day doesn't need a person which means it's just software anyone can create

it's just software anyone can create that software software creation is so

that software software creation is so cheap right now so as long as you have a

cheap right now so as long as you have a verifiable test for the agent and it's

verifiable test for the agent and it's open source which react doctor is it's

open source which react doctor is it's it's a commodity this that that loop is

it's a commodity this that that loop is a commodity you cannot build a product

a commodity you cannot build a product out of it because you will get margin to

out of it because you will get margin to zero but let's say you have a custom

zero but let's say you have a custom test custom verifier that no one else

test custom verifier that no one else has been able to build out and you have

has been able to build out and you have a closed loop system no one else has you

a closed loop system no one else has you actually have a very profitable business

actually have a very profitable business cuz for you it's fully automated for

cuz for you it's fully automated for others it's not and no one knows how you

others it's not and no one knows how you are closing that loop another one is

are closing that loop another one is just logistics operations business most

just logistics operations business most of it is automated a part of it is done

of it is automated a part of it is done by humans when agent does need that

by humans when agent does need that feedback and that whole system works in

feedback and that whole system works in an economy of scale and this is a real

an economy of scale and this is a real that is a really good example. I worked

that is a really good example. I worked at a company called Fathom and before

at a company called Fathom and before LLMs were a thing transformers were

LLMs were a thing transformers were still a thing. I mean LLMs were still a

still a thing. I mean LLMs were still a thing but they were BERT models. It

thing but they were BERT models. It wasn't GPT model. So we were training a

wasn't GPT model. So we were training a bird model to do medical coding. And

bird model to do medical coding. And medical coding basically is you go to a

medical coding basically is you go to a doctor you explain everything they're

doctor you explain everything they're usually typing or now they are using

usually typing or now they are using scribes to record everything. They

scribes to record everything. They that's called an encounter. You

that's called an encounter. You encountered with the doctor doctor wrote

encountered with the doctor doctor wrote down what you are experiencing all that

down what you are experiencing all that that encounter then converts to

that encounter then converts to diagnosis and prognosis and then those

diagnosis and prognosis and then those goes to the insurance company actually

goes to the insurance company actually those gets converted to medical codes

those gets converted to medical codes and then goes to insurance company for

and then goes to insurance company for billing. So then insurance company has

billing. So then insurance company has medical coders that will verify that

medical coders that will verify that okay because of this diagnosis this is

okay because of this diagnosis this is the amount of billing that they're

the amount of billing that they're asking and it matches the actual

asking and it matches the actual encounter notes and the patient's notes

encounter notes and the patient's notes and that's like a whole process. What

and that's like a whole process. What Fathom was trying to do was automate it

Fathom was trying to do was automate it using language models. So you throw the

using language models. So you throw the text raw text of the encounter here. You

text raw text of the encounter here. You get the medical codes out here and it's

get the medical codes out here and it's as accurate as possible. And the way we

as accurate as possible. And the way we we implemented we went to market with

we implemented we went to market with this initially was we didn't have the

this initially was we didn't have the system working. So we went to like when

system working. So we went to like when customer was interested in in doing this

customer was interested in in doing this they were sending out their volume to

they were sending out their volume to overseas company. Fathom just said hey

overseas company. Fathom just said hey we will do this at like 50% the cost.

we will do this at like 50% the cost. And they they took the cost, they had

And they they took the cost, they had the labelers label, get that data, they

the labelers label, get that data, they created that data from nothing. And now

created that data from nothing. And now that they had the data, they started

that they had the data, they started training the model. And now 80% 85% even

training the model. And now 80% 85% even 90% of those records are automated. And

90% of those records are automated. And 10% are still going to people because

10% are still going to people because the model is saying a score is not high

the model is saying a score is not high enough. And this is a very traditional

enough. And this is a very traditional like machine learning. It's not using

like machine learning. It's not using LLMs like today. It they build big

LLMs like today. It they build big models like foundational models pretty

models like foundational models pretty much for this. And it's a it's a it's a

much for this. And it's a it's a it's a very profitable company running today.

very profitable company running today. Um you know series B they raise series B

Um you know series B they raise series B they are probably going to raise series

they are probably going to raise series C soon. And yeah it's it's just like

C soon. And yeah it's it's just like those kind of systems is how I see the

those kind of systems is how I see the world is going to move is you will

world is going to move is you will either own a workflow where verification

either own a workflow where verification is very hard and you can build out a

is very hard and you can build out a team and and agents are doing most of it

team and and agents are doing most of it and some people are coming in to fill in

and some people are coming in to fill in that gap because at the end of the day

that gap because at the end of the day that's the easiest way to get a service.

that's the easiest way to get a service. If I want something done, I don't care

If I want something done, I don't care if agent is doing it or I'm doing it

if agent is doing it or I'm doing it with my agent. I just care if it's done

with my agent. I just care if it's done or not and I if I can send it out to you

or not and I if I can send it out to you and then you can give me the result

and then you can give me the result back. It's more like a BO type of uh

back. It's more like a BO type of uh model for those kind of things. And then

model for those kind of things. And then other things could be I want my agent to

other things could be I want my agent to do it. And that's more like yeah me

do it. And that's more like yeah me setting up my agent and and and because

setting up my agent and and and because I want control over data. I want control

I want control over data. I want control over things. I don't want to give you

over things. I don't want to give you that access to my data.

that access to my data. Right. Right. Right. looking just just

Right. Right. Right. looking just just to switch gears for a second. Looking

to switch gears for a second. Looking into the future, what did you personally

into the future, what did you personally believe about agents say 12 months ago

believe about agents say 12 months ago that you don't believe any longer?

that you don't believe any longer? Very good question. These kind of

Very good question. These kind of questions always catch me off guard cuz

questions always catch me off guard cuz but I'll think about it. I The reason

but I'll think about it. I The reason they catch me off guard is cuz I never

they catch me off guard is cuz I never think like this. It's like, oh, what did

think like this. It's like, oh, what did this happen that I never happened? It

this happen that I never happened? It was not the funniest.

was not the funniest. That's fair. But uh but this is a good

That's fair. But uh but this is a good question because it it does make you

question because it it does make you reflect on okay what were your priors a

reflect on okay what were your priors a year ago. I think it's very little that

year ago. I think it's very little that I don't believe. I think the only thing

I don't believe. I think the only thing that has changed is actually I I now

that has changed is actually I I now have way more trust of or I would say I

have way more trust of or I would say I think it came way faster than I thought

think it came way faster than I thought it was going to be.

it was going to be. Okay. I still believe everything is

Okay. I still believe everything is going to be possible but I thought by

going to be possible but I thought by end of this year we would be at this

end of this year we would be at this stage and we already passed that in

stage and we already passed that in March of this year right so let's say

March of this year right so let's say January of la this year I had some

January of la this year I had some predictions I thought by end of this

predictions I thought by end of this year we would at least have you know

year we would at least have you know this dynamic workflow thing xyz and this

this dynamic workflow thing xyz and this is how the agents would do it and I can

is how the agents would do it and I can see it and then claude released dynamic

see it and then claude released dynamic workflow around March sometime and I was

workflow around March sometime and I was just shocked that wow this came so much

just shocked that wow this came so much faster than I thought and it actually

faster than I thought and it actually works because my predictions are mostly

works because my predictions are mostly based on how fast the model improves and

based on how fast the model improves and so far I have not been able to catch

so far I have not been able to catch that exponential mentally even though I

that exponential mentally even though I I projected that okay but I always

I projected that okay but I always projected from the past experience and

projected from the past experience and and humans we are very bad at the

and humans we are very bad at the exponential so the only thing that

exponential so the only thing that surprises me is like my predictions on

surprises me is like my predictions on how fast models are improving is not is

how fast models are improving is not is not accurate for some reason

not accurate for some reason I I mean I think you're joining every

I I mean I think you're joining every single other human being on earth

single other human being on earth I think like when you yeah every

I think like when you yeah every interview you hear from the frontier

interview you hear from the frontier labs the leaders are shocked at the

labs the leaders are shocked at the progression. So I think that's very

progression. So I think that's very fair. Don't be too hard on yourself on

fair. Don't be too hard on yourself on that prediction. I don't think anyone's

that prediction. I don't think anyone's getting

getting predictions. Everyone knows how the how

predictions. Everyone knows how the how it will turn out and you're like, "Okay,

it will turn out and you're like, "Okay, by this time we would have this and and

by this time we would have this and and then it comes early and it's it's it's a

then it comes early and it's it's it's a certifi is a good it's a good thing,

certifi is a good it's a good thing, right? Like it's a good that it's

right? Like it's a good that it's faster.

faster. Yeah. Yeah. Yeah. Yeah. And and so just

Yeah. Yeah. Yeah. Yeah. And and so just moving if it's okay to wrap up, I want

moving if it's okay to wrap up, I want to ask just a few quick questions, quick

to ask just a few quick questions, quick answers. One question I had for you is

answers. One question I had for you is what business leader or or builder

what business leader or or builder influences you and how you build out a

influences you and how you build out a work worker or just how you operate

work worker or just how you operate day-to-day.

day-to-day. I think the northstar is always Steve

I think the northstar is always Steve Jobs in terms of

Jobs in terms of Okay, nice.

Okay, nice. Like I I still listen to a lot of Steve

Like I I still listen to a lot of Steve Jobs old interviews. I would have like

Jobs old interviews. I would have like if you go to my YouTube history, some of

if you go to my YouTube history, some of the interviews would be over 100 times.

the interviews would be over 100 times. That's just because there was so much

That's just because there was so much profound insights in simple sentences

profound insights in simple sentences that were captured. But recently, I

that were captured. But recently, I would also say Travis Kalanick from Uber

would also say Travis Kalanick from Uber has been more public recently and I have

has been more public recently and I have watched every single interview of his as

watched every single interview of his as well. And it's always very inspiring to

well. And it's always very inspiring to listen to him because after you listen

listen to him because after you listen to him, you realize that oh wow, the

to him, you realize that oh wow, the world is not as restricted as you

world is not as restricted as you thought it was. Like you can really take

thought it was. Like you can really take big swings. So recently I would say I've

big swings. So recently I would say I've been enjoying his sort of like

been enjoying his sort of like interviews that he's been doing. He did

interviews that he's been doing. He did one with TBPN. He did one with uh E16Z.

one with TBPN. He did one with uh E16Z. Listen to both of them. They were great.

Listen to both of them. They were great. So I guess the common thread there is

So I guess the common thread there is both of them are contrarians like they

both of them are contrarians like they think against the grain. Both Jobs and

think against the grain. Both Jobs and Travis.

Travis. Yeah. I feel like Jobs has so Steve Jobs

Yeah. I feel like Jobs has so Steve Jobs is just it's like so fundamentally

is just it's like so fundamentally human, right? It's

human, right? It's right. You start with the human

right. You start with the human experience and build back towards

experience and build back towards technology and and these are lessons

technology and and these are lessons that technologists and engineers like

that technologists and engineers like myself forget often and we are like we

myself forget often and we are like we look at the agent what it can do and

look at the agent what it can do and then we try to take it to market and it

then we try to take it to market and it should always be and I think when you go

should always be and I think when you go reverse I think that's when you innovate

reverse I think that's when you innovate because if human wants this experience

because if human wants this experience and that's not available you go and

and that's not available you go and figure it out. So 2 years ago if

figure it out. So 2 years ago if companies were doing voice AI they were

companies were doing voice AI they were just looking at what models were

just looking at what models were available and they did that versus GPT

available and they did that versus GPT has been pushing boundaries some other

has been pushing boundaries some other voice companies or AI labs have been

voice companies or AI labs have been pushing boundaries of duplex voice mode

pushing boundaries of duplex voice mode and without that if you're just an

and without that if you're just an engineer building on existing models not

engineer building on existing models not going to work. So if you think from

going to work. So if you think from Steve Jobs's lens your ambition becomes

Steve Jobs's lens your ambition becomes bigger that okay the duplex mode is not

bigger that okay the duplex mode is not on as in if I talk agent stops and if

on as in if I talk agent stops and if agent talks I have to stop and etc and I

agent talks I have to stop and etc and I have to interrupt. That's not the way

have to interrupt. That's not the way human conversation works. And if you

human conversation works. And if you just take that perspective, your goal

just take that perspective, your goal then becomes, okay, let me hire

then becomes, okay, let me hire researchers who can actually do this.

researchers who can actually do this. Let me go raise more money. Makes your

Let me go raise more money. Makes your ambition bigger. Um, and same with

ambition bigger. Um, and same with Travis as well. It's

Travis as well. It's he always looks at the final state of

he always looks at the final state of how the system how the world should be.

how the system how the world should be. His his real model right now is that

His his real model right now is that there's physical computer, there's a

there's physical computer, there's a bits computer, which is what we use

bits computer, which is what we use today, but there's also Adams computer

today, but there's also Adams computer and that's what he's trying to build,

and that's what he's trying to build, which is this company Adams. That was

which is this company Adams. That was also very interesting for me because he

also very interesting for me because he looks at the world as his final state

looks at the world as his final state and he's trying to build the hard

and he's trying to build the hard technology to get to that stage. Not

technology to get to that stage. Not just what's available in the market, let

just what's available in the market, let me let me build a product around it and

me let me build a product around it and let me sell it. That's literally 99% of

let me sell it. That's literally 99% of the businesses in the world. And the 1%

the businesses in the world. And the 1% that innovate, they go back from like

that innovate, they go back from like the end state to to what what they need

the end state to to what what they need to do.

to do. Yeah. Yeah. And I think having proof of

Yeah. Yeah. And I think having proof of doing it once before helps a lot on this

doing it once before helps a lot on this current venture he's moving forward

current venture he's moving forward with. And and just for for yourself,

with. And and just for for yourself, what do you do to quiet your mind

what do you do to quiet your mind outside of building work layer?

outside of building work layer? Yeah, meditation is mostly what I do. I

Yeah, meditation is mostly what I do. I have been doing it less and less, but I

have been doing it less and less, but I do want to get back into it. It's the

do want to get back into it. It's the best way I've found peace when a lot of

best way I've found peace when a lot of things are happening.

things are happening. Yeah. So, do are you not at peace now

Yeah. So, do are you not at peace now moving away from meditation? Do you feel

moving away from meditation? Do you feel like

like No, I feel like my mind's running in

No, I feel like my mind's running in like hundred other directions all the

like hundred other directions all the time.

time. And

And it's so energizing to keep let it run.

it's so energizing to keep let it run. I And I think this is one of the things

I And I think this is one of the things where you don't know how much you miss

where you don't know how much you miss your piece until you meditate and and

your piece until you meditate and and find it. So right now it's just it's

find it. So right now it's just it's just like oh this thing we need to

just like oh this thing we need to launch. We have this new product coming

launch. We have this new product coming out

out and you wake up in the morning and

and you wake up in the morning and you're like I have 100 things to do. let

you're like I have 100 things to do. let me just go do them instead of like sit

me just go do them instead of like sit down and meditate. Sometimes it's it's

down and meditate. Sometimes it's it's beneficial to take that hour out because

beneficial to take that hour out because you might clear your thoughts and and

you might clear your thoughts and and that's why I want to do it more. It's

that's why I want to do it more. It's not because of the peace, but it's

not because of the peace, but it's because maybe I'm missing something and

because maybe I'm missing something and I have 100 things and maybe I could cut

I have 100 things and maybe I could cut 90 of them and maybe meditation

90 of them and maybe meditation and that calm and peace will tell me

and that calm and peace will tell me which 90 to cut.

which 90 to cut. Yeah.

Yeah. That's where that's where it really

That's where that's where it really comes in handy. So, it saves time by

comes in handy. So, it saves time by like taking your hour basically.

like taking your hour basically. Yeah. Yeah. Like it Yeah. It it allows

Yeah. Yeah. Like it Yeah. It it allows your brain to process in a different

your brain to process in a different way. I hope you get to get back to it. I

way. I hope you get to get back to it. I know you're launching a new product, so

know you're launching a new product, so it's probably quite chaotic, but lastly,

it's probably quite chaotic, but lastly, just before I let you go, what does

just before I let you go, what does success look like for you personally?

success look like for you personally? Like what does it ultimately look like?

Like what does it ultimately look like? Yeah. So, for me, what I really am

Yeah. So, for me, what I really am obsessed about right now is getting

obsessed about right now is getting agents to do work that we don't really

agents to do work that we don't really like doing. And my work is already like

like doing. And my work is already like that. I am not doing a lot of the works

that. I am not doing a lot of the works that I hated. I am doing a lot of

that I hated. I am doing a lot of strategic thinking, sitting in front of

strategic thinking, sitting in front of likeund at this point. I probably have

likeund at this point. I probably have 20 40 sessions open. That doesn't mean I

20 40 sessions open. That doesn't mean I I have agent running in every single one

I have agent running in every single one of them. It's just 20 40 different

of them. It's just 20 40 different threads that are doing different work

threads that are doing different work open and then I can go and and inject my

open and then I can go and and inject my thoughts into different ones whenever I

thoughts into different ones whenever I have time and some have high priorities

have time and some have high priorities so they are the first one where I inject

so they are the first one where I inject my thought other ones can be like they

my thought other ones can be like they are still for a seven seven days right

are still for a seven seven days right but this is the high strategy work that

but this is the high strategy work that I'm doing where I can have 40 threads

I'm doing where I can have 40 threads open and every single thing is recorded

open and every single thing is recorded and I'm just talking I think this can be

and I'm just talking I think this can be a lot of people's job right you you go

a lot of people's job right you you go in and you just talk to your computer

in and you just talk to your computer and actions are being done behind the

and actions are being done behind the scenes for you You're not doing the

scenes for you You're not doing the manual [clears throat] drag and drop and

manual [clears throat] drag and drop and all that. You're not changing the words

all that. You're not changing the words in spreadsheet. You're just reviewing.

in spreadsheet. You're just reviewing. You're giving your opinions. This is

You're giving your opinions. This is what CEOs and and leaders today do

what CEOs and and leaders today do anyways. They have assistants. They have

anyways. They have assistants. They have people hired to do that. I think we can

people hired to do that. I think we can we can bring that to every single job

we can bring that to every single job which means you basically take human

which means you basically take human from doing manual labor or even manual

from doing manual labor or even manual labor on computer is just like clicking

labor on computer is just like clicking and doing all that to just thinking just

and doing all that to just thinking just going from your brain to to results.

going from your brain to to results. That's what we are obsessed about. So

That's what we are obsessed about. So for us success is doing that but the

for us success is doing that but the best way to measure it is how much of

best way to measure it is how much of like let's say token consumption today

like let's say token consumption today comes from like certain products cloud

comes from like certain products cloud code cursor codeex a bunch of products

code cursor codeex a bunch of products have like the biggest chunk of it open

have like the biggest chunk of it open code is also like a very big piece now

code is also like a very big piece now and you can think of the inference like

and you can think of the inference like the GPUs running are coming through some

the GPUs running are coming through some product surface and then going to the

product surface and then going to the GPU right they're not like just people

GPU right they're not like just people using APIs themselves there's only

using APIs themselves there's only developers who are doing that they're

developers who are doing that they're building a product and then still end

building a product and then still end users are doing most of it. So our goal

users are doing most of it. So our goal is that we become a product surface that

is that we become a product surface that drives up the GPU inference because a

drives up the GPU inference because a lot of buildout is happening behind the

lot of buildout is happening behind the scenes and at the end of the day we just

scenes and at the end of the day we just want to drive that GPU usage up more

want to drive that GPU usage up more intelligence that people want to use.

intelligence that people want to use. They can they can find that product

They can they can find that product surface through us the same way they

surface through us the same way they found through cursor. the same way same

found through cursor. the same way same way they're found through codecs and

way they're found through codecs and cloud code when those things came out

cloud code when those things came out and it just took so our new product

and it just took so our new product launch is basically that we are looking

launch is basically that we are looking ways and our success metric is that at

ways and our success metric is that at the end of the day we have so much usage

the end of the day we have so much usage passing through us that we have to buy

passing through us that we have to buy our own GPU data center right um and

our own GPU data center right um and then we have to serve our own for our

then we have to serve our own for our own customers so that's like the real

own customers so that's like the real real success

real success either we we partner with SpaceX and

either we we partner with SpaceX and take like you know have a big contract

take like you know have a big contract with them to have like five dedicated

with them to have like five dedicated like data centers for us or

like data centers for us or it's like something that we just have to

it's like something that we just have to physically invest in either on land or

physically invest in either on land or on space and and get some company to

on space and and get some company to build it out for us.

build it out for us. Wow. And this Yeah, I mean that is that

Wow. And this Yeah, I mean that is that is a a a well articulated ambitious goal

is a a a well articulated ambitious goal for the agents that you're you're or the

for the agents that you're you're or the future you're building out for humans to

future you're building out for humans to work very closely with agents. So

work very closely with agents. So yeah, and it all comes with the product

yeah, and it all comes with the product layer. If the product layer is good,

layer. If the product layer is good, which is I think cursor actually

which is I think cursor actually accomplished the goal that I want to do.

accomplished the goal that I want to do. They build a product so good and now

They build a product so good and now they are partnered basically they got

they are partnered basically they got acquired by SpaceX and now they're going

acquired by SpaceX and now they're going to even increase the usage of agents

to even increase the usage of agents even more. Right.

even more. Right. Sure.

Sure. So there are companies that are already

So there are companies that are already sort of like proven out this concept

sort of like proven out this concept that you can start from the product

that you can start from the product layer and then you can move up to like

layer and then you can move up to like the inference layer.

the inference layer. Basically that's what we want to do.

Basically that's what we want to do. Yeah. I I mean I wish you all the best

Yeah. I I mean I wish you all the best on your pursuit. I mean, as someone

on your pursuit. I mean, as someone who's met you before this podcast and

who's met you before this podcast and and plans to use your products, I think

and plans to use your products, I think it's awesome what you're working on. So,

it's awesome what you're working on. So, we can we'll wrap there. Uh, this is

we can we'll wrap there. Uh, this is really good. I really enjoyed speaking

really good. I really enjoyed speaking with you, Mel. Thank you for joining us

with you, Mel. Thank you for joining us today.

today. It was a great great chat. Behind

It was a great great chat. Behind [music] the throw.

[music] the throw. Behind the door.

Behind the door. One move and the whole thing goes.

One move and the whole thing goes. [music]

[music] Yeah. We build. We build. We build it

Yeah. We build. We build. We build it real [music] behind the