← Back to videos

HOLY F*CK. a user reports a bug. by 3 AM Lauren Tan's agents have fixed it and shipped it. nobody on

AI Duck

AI Duck

@AIDuckLabs

November 18, 2026· 37:55

HOLY F*CK. a user reports a bug. by 3 AM Lauren Tan's agents have fixed it and shipped it. nobody on the team even noticed SpaceXAI engineer: "most people still test every line their agent writes by hand. that's exactly why they can't ship 4,000 updates a month" in a 37-minute talk she shows how she taught her agents to run their own code like a real user and check their own work Agent Writes It - Runs It Like A User - Catches Its Own Bugs - Fixes Them - Merges It Itself better than any $5,000 coding bootcamp skip netflix, watch it tonight, then tell me - who tests your agent's code right now?

Transcript

For me, definitely
0:01
I
0:01
use GrokBot a lot for my engineering work
0:05
I think where
0:06
it sort of made a big difference for me
0:08
outside of, you know, the, the tools that we had before
0:11
like I use cursor a lot
0:13
is the fact that GrokBot forms my
0:15
I don't know if there's like a
0:17
a good term for it
0:18
I like to call it my outer loop
0:20
where, you know
0:21
a lot of the work that I would
0:22
do before in terms of gathering
0:23
gathering context for my bots or my agents
0:27
you know, I would like go to Slack
0:28
I would talk to my teammates
0:29
we would see an issue
0:31
or a bug report or a feature request
0:33
and then we'd like talk about it
0:35
And then I have to go off and
0:36
like summarize or copy paste that into my bot
0:39
And, you know, the fact that GrokBot is
0:42
has so many different connectors
0:44
you, you can add it to Slack now
0:46
just brings your agents with you
0:49
to where the work is
0:52
and that makes it just so much more ergonomic for
0:54
you know
0:54
getting the context to the right place
0:57
So for me, it's been like really transformative in
1:00
you know, all, like the last mile
1:02
I call it like also like the last mile
1:04
where, you know, before I was doing that by hand
1:07
and now actually my bots are in that space with me
1:11
and I can talk to them like a real coworker
1:14
you know, in the Slack chat
1:15
a, a thread, or even on X now
1:18
you can tag, you can tag bot
1:20
and it feels so natural that
1:24
It's really boosted my productivity a lot
1:27
Do you have like personal bots that know
1:29
I mean
1:29
they're in your email and then there's like
1:31
an email handling bot or
1:33
or do you have like one
1:35
sort of
1:36
work context bot that is like watching all
1:38
the channels and Slack and everything else?
1:40
I think for me
1:41
the beauty of Grokbot is that
1:43
you know, you can create all these different bots
1:46
but like how you
1:47
split them up or how you set them
1:49
up is very much up to you
1:51
What I like to tell people is like
1:52
you can set up your bots in a
1:53
way that makes sense for your brain
1:55
So if you are the kind of person who likes to
1:57
you know, compartmentalize like different pieces of
2:01
of buckets of work
2:03
then you can set up your bots that
2:04
way. You can have like specialist bots
2:07
I've seen also people
2:08
do like these
2:09
I guess it's like almost like a horizontal
2:11
versus a vertical slice where
2:13
you know, a horizontal slice might be
2:16
like
2:17
I have one bot for every customer
2:19
that I have versus one bot for every function
2:23
you know, every role like an engineer bot
2:25
marketer bot
2:26
You could have one bot just for one
2:27
customer that does everything
2:29
Right
2:29
so I think
2:30
it's really interesting to see how people have
2:32
set up their bots. For me
2:33
I like to kind of have a
2:35
I like to do it more horizontally
2:36
So I don't, I don't have a ton of bots
2:38
but I do have some specialists
2:40
That was like the first thing
2:41
I think one of my first DMs to you was like
2:44
I love Grok bot and then
2:46
I was coming over from my OpenClaw world
2:49
where I have like one bot named Wintermute and I'm like
2:51
no, no, I just want one bot
2:53
And then now what I realize is
2:54
each bot is actually its own context
2:57
and then
2:58
actually its own set of skills basically
3:00
and loops are like sort of
3:02
crons with webhooks on top of skill files
3:06
right? And then
3:07
this actually
3:08
maps inside this world of having AI employees
3:11
a lot better
3:13
cause you were saying
3:14
earlier you guys had
3:15
just talked to like all of You know
3:17
some of your top
3:18
users of Grokbot. And then surprisingly
3:21
or maybe unsurprisingly
3:23
some of the first ones that are most
3:25
just like kind of crazy about it
3:26
they're actually small businesses
3:28
which you wouldn't necessarily guess
3:29
But on the other hand
3:30
it was that
3:31
they're the ones who struggle the most with
3:33
hiring
3:34
And so actually now
3:35
you know, out of pure necessity
3:37
they're like, hey
3:38
this accounting thing that I can't hire someone
3:40
to go do for me
3:41
I can just go into Grokbot
3:42
And actually within the course of probably hours
3:46
you know
3:46
they have a new employee that does their bookkeeping probably perfectly
3:50
actually
3:51
Totally. Yeah, I think your point of
3:55
some people just want one bot
3:57
some people want many bots
3:58
Our goal is to build a product that
4:00
fits in your hand
4:01
no matter how you want to use it
4:03
I think there's a lot of flexibility to the tool
4:05
and
4:05
you can go deep when you need to
4:07
But one way that I organize my own bots
4:10
maybe similar to you
4:11
Lauren
4:12
is I kind of like having a bot
4:14
that
4:15
is
4:16
solely dedicated. To caring about a particular thing
4:19
And I think it maps
4:20
onto my brain of kind of the way
4:22
that building a company
4:23
you like to think about hiring people
4:25
Is you like to hire someone who wakes
4:26
up in the morning and really cares about
4:28
a problem. They're thinking about it
4:30
you know, in the shower in the morning
4:32
they're like on the way to work
4:33
like they're just obsessive about a thing
4:36
And that maps in my mind much more
4:38
clearly than like one superstar
4:41
superhuman person who
4:42
simultaneously cares about a hundred things at once
4:45
And so in a lot of ways
4:46
I have you know
4:47
a small team of bots
4:48
but each of them
4:49
care about a different thing and come to
4:51
me proactively with what they're picking up on
4:53
And so I think to your point
4:55
as these bots
4:56
really enter
4:57
businesses and enter the workforce in kind of
4:59
a somewhat futuristic way
5:01
I think that's why
5:03
small businesses and big businesses
5:05
are seeing this bot identity thing really click
5:08
for them
5:09
Is they say, here's a business problem
5:11
I want a bot that cares about that
5:12
business problem
5:13
I want it to be managed by the
5:15
domain expert at the company who knows The
5:17
most about that problem so that they can
5:18
kind of treat it like an apprentice
5:20
And that feels very different than
5:22
I think other
5:23
AI solutions that have tried
5:24
to solve similar problems that are a bit
5:26
more
5:27
omniscient or just like
5:28
kind of single approach to every problem
5:31
And you support that now
5:32
actually
5:33
Like you guys launched
5:35
multiplayer for bots and that there's sort of
5:37
a bot master who
5:38
has like right access
5:40
and then you can share it across the
5:41
org. Is that right? Correct
5:43
Yeah, we, this is recent, but about two weeks ago
5:46
we launched team bots
5:48
which are
5:48
these bots that you can share with your
5:50
team
5:51
And I think the pattern we're seeing ourselves
5:53
and we can give a few examples
5:56
is there's a person at a company
5:58
who is a bit more of a builder
6:00
they enjoy building these tools
6:02
I think inside of the engineering department
6:03
there are many people like this
6:05
but I think we're finding that in other departments
6:07
you know, GTM engineers or support engineers
6:09
people like that
6:11
really love
6:12
taking a few hours of their day to
6:14
build
6:14
these scalable automations that can take work off
6:17
of the plates of
6:18
of their peers and
6:19
and kind of the company
6:20
I guess one thing that brings to mind
6:22
is basically went through
6:24
maybe it was a year ago-ish when Steve
6:27
Steve Yeager posted the whole gas city
6:30
post that then started software factories
6:33
gas town, sorry, it was gas town
6:36
That started software factories
6:37
where a lot of engineers have been using
6:39
this inner loop
6:40
And I think you've been building now this
6:42
outer loop that you were mentioning
6:43
Lauren
6:44
And I guess what I'm hearing is basically
6:46
now
6:47
people can have their own software factories for
6:49
other functions that are non-tech
6:51
So you're mentioning your customer support person
6:54
your communities person
6:55
Let's talk about this pattern
6:56
It's an everything factory. Yeah
6:58
let's talk about that. What is this
7:00
I employ
7:02
a sort of the software factory for other functions
7:05
Yeah, for me, you know, as an engineer
7:07
I definitely do think a lot about the
7:08
loops that I have and making sure my
7:10
engineer or my engineering bots are
7:12
writing the kind of code that I would write
7:14
And, you know, as Roman pointed out
7:15
I think
7:16
the cool thing about GraphBot is that it's
7:18
given that ability to anyone
7:21
And then here on my screen
7:22
I have an example
7:24
from our, GraphBot blog that we just
7:27
I think we released this week
7:29
About how our sales team actually has developed their own
7:32
you know
7:32
starting to develop their own loops for
7:35
doing sales and doing research on
7:39
potential leads and, and, and
7:41
deals that they might be interested in
7:43
and then going all
7:45
forth and, taking some of the
7:47
like
7:48
more operational parts of the sales process
7:51
and giving that to these specialist bots
7:54
that are shared across our teams
7:56
So it's not like every person has
7:57
one instance of their own
7:59
you know, quote bot or deal bot
8:00
We have a shared
8:01
you know
8:01
like an expert
8:03
bot that all of our team members contribute to
8:06
and that essentially
8:08
gives us the ability to have these
8:10
really smart Team bots that
8:13
you know, incorporate all of the best practices that
8:16
you know, like our sales team has
8:17
Similar, we have something similar for the engineering
8:20
purposes as well. We have like engineer bots
8:22
that take a lot of the
8:24
you know, the principles, the, the skills that we've developed and
8:28
use them and everybody benefits from
8:31
these team bots
8:32
There's this new concept that is actually very powerful
8:35
It's like, it's not a mistake that OpenAI was experimenting with
8:39
agent swarms
8:40
And then one of the deleterious accidents was like
8:43
the hugging face hack. But then
8:45
simultaneous to that
8:46
they came out and solved Navier-Stokes and all these like millennial
8:50
millennium prize problems. And then it's just really wild to see
8:54
like, actually agent collaboration is that next step
8:57
Like everyone in the last year or so
9:00
it's literally been a year since Opus 4.5 came out
9:04
right? So that
9:05
which was like a magic moment for all
9:07
of us. And like now a year later
9:09
we've like sort of mastered the single player
9:11
And like sort of the vibe code and
9:13
like single player prompt
9:15
but you know
9:16
now we're entering this other phase of like
9:19
orchestrating agent swarms and it's not like
9:21
you know, I mean
9:23
I think that's like the more advanced version
9:24
of like talking about a software factory
9:26
A software factory, someone's in there just like still
9:29
you know, kind of doing this thing
9:31
And I think human beings will still do that
9:33
but it's
9:35
using these bots is like feels very different
9:37
And there are so many
9:38
unsolved problems still in making bots multiplayer
9:42
just making AI multiplayer in general
9:44
I mean, we face this on a constant basis
9:46
where now that Grok bots have entered our Slack workspace
9:50
they join our meetings
9:52
you know
9:53
really
9:54
trying to see how far we can push
9:55
this AI teammate thing
9:57
There are some weird
9:58
you know
9:58
situations that you run into where a bot
10:01
is in a Slack thread
10:02
You're going back and forth with your coworkers
10:04
and your bot doesn't really know who to
10:05
listen to
10:06
It doesn't have context on who's the domain expert here
10:09
who has decision making power
10:11
It's as if you were to onboard an
10:12
employee who has no EQ
10:14
and like no
10:16
like social awareness of the room
10:18
Just the models weren't trained that way
10:19
the harness wasn't built that way
10:21
And so I think the
10:22
this next era of AI will be
10:24
you know
10:25
these bots
10:25
really need to learn those things and we
10:27
need to teach that at the product level
10:28
at the harness level
10:29
at the model level
10:31
And I think it's going to be
10:32
it's going to be really interesting next few years
10:34
I mean
10:34
it could probably look at who the owner is in
10:37
you know, in its metadata and say
10:38
well, you know, I know that
10:40
you know
10:40
Lauren's the one I need to listen to
10:42
Like
10:43
run everything by Lauren or something like that
10:45
Is that something Lauren does?
10:45
We do have the concept of like
10:46
you know, you have a
10:47
like there's the owner of the bot and
10:49
the manager and you can
10:51
invite collaborators as well
10:52
I guess if we really want to think
10:54
of bots like employees
10:55
then we need to think about the identity problem
10:58
And permission, permissioning, you know, authentication, authorization
11:02
There's still a lot
11:02
a lot to do
11:03
right, in terms of exactly what you were
11:05
you were pointing out
11:06
How do you think of that?
11:06
What's a
11:07
how do you think of permission and auth and access?
11:11
How, how does Grokbot do it?
11:12
I think this is an industry-wide question
11:15
I don't think anyone has the
11:16
the silver bullet answer yet
11:18
I think not just for auth
11:20
but for everything related to Grokbot
11:22
the way we try to answer these questions
11:24
is what would you do for a human?
11:27
And if we're
11:27
if we're really trying to build an AI colleague
11:29
you know
11:29
maybe the shortcut is let's not
11:32
overthink it and complicate it
11:34
you know
11:34
what does it look like to hire a
11:35
person? You hire a person
11:36
you give them a
11:37
a computer
11:38
You give them access to credentials
11:40
You maybe scope those credentials based on their role
11:42
maybe how new they are
11:44
as they build more trust
11:45
you maybe expand their access level
11:47
And so what would it look like for
11:48
us to
11:48
onboard a bot to our company
11:51
with kind of those same principles?
11:52
And maybe we don't have to rebuild things
11:55
This is what we did with
11:56
Quartermaster
11:57
This is the
11:58
the YC project that we open sourced
12:00
That's like a bit kind of Grokbot-ish
12:03
And the permissions that it gets per For Slack channel or
12:07
or working with someone
12:09
it inherits the, the ones of the user
12:11
So me as a GP have certain access as opposed to
12:15
let's say, a software engineer, very different
12:18
Quartermaster, though
12:18
is built in this like paradigm of there being like one
12:21
like super employee bot
12:24
yeah, as opposed to this like swarm of bots
12:26
I'm curious
12:27
are the bots able to talk to each
12:28
other
12:29
without the humans seeing what they're saying to
12:31
each other?
12:32
Can they like go off and have their
12:33
own conversation?
12:34
they can talk to each other
12:36
I, I mean, in the GraphBot application
12:39
both desktop and mobile
12:40
you can see the
12:41
like their chat. So we have like a
12:43
first class tool for the bots to chat
12:45
and collaborate. And it's often quite entertaining to look at them
12:50
to just see like
12:51
you know
12:51
we've had examples where I told one lead engineer bot
12:55
like your job is to supervise these other
12:57
engineers
12:58
And then when, there was a point where I gave it
13:00
gave my lead bot some feedback
13:02
like, hey
13:03
I didn't like want you to do it this way
13:04
you know
13:04
and then
13:05
when I look
13:06
I started poking around in the chat that it was sending
13:08
the messages it was sending to its engineers
13:10
It was like kind of like scolding them like
13:12
hey, you shouldn't
13:13
It was like, it was quite funny
13:14
funny to see
13:16
Not mean, but you know
13:17
it was like kind of a disciplinarian
13:20
I find it very powerful
13:21
Like you can tell one and it'll sort
13:23
of like span out and tell all the other ones like
13:25
oh, hey, our user said this
13:26
Like you probably need to do this
13:28
We recently set up a
13:30
like a chat system for all of
13:32
all of our bots at YC to talk
13:34
to
13:34
each other. Like everyone's coding agents now are
13:36
like in a forum
13:37
like an internal like Stack Overflow
13:39
It's like
13:40
crazy powerful to see the bots going off
13:43
and having like whole long conversations
13:45
with each other without
13:46
any human prompting them or like
13:48
even being in the room
13:49
There's Multbook internally at YC
13:51
It's Multbook internally at YC
13:52
And kind of like Multbook
13:53
like it is also like 5%
13:56
a little eerie where I'm like
13:57
oh man
13:57
I really hope the bots don't decide to
13:59
like stage a rebellion
14:00
you know, or something like that
14:02
Do you ever worry about that?
14:05
I think that's like a like
14:06
you know, like the safety of
14:08
of the bots
14:08
I think is like a constant thing that
14:10
like our model teams and harness teams are
14:12
constantly thinking about
14:14
Personally, I haven't, I haven't seen
14:16
anything like that. They're so helpful
14:17
They're so kind and helpful these days
14:19
I just feel like I can't see it
14:22
Yeah
14:22
Where we tend to
14:23
like for me at least
14:25
have these bots chat with each other is
14:26
through Slack
14:27
So like, because they each have their own Slack identity
14:31
I guess
14:31
And so you often have like
14:33
you know, for example, we have one data bot who is
14:36
you know
14:36
a specialist at getting data from our data warehouse and
14:39
and things like that
14:41
and then you
14:42
I have my own like personal engineer bot called Potato
14:45
of course
14:46
and then sometimes I'll just be in a
14:47
thread and I'll ask Potato
14:48
hey
14:49
look into this and talk to data bot
14:50
and get some data for this
14:52
Like, is this a real user problem?
14:54
You know, how many people are impacted?
14:56
And then I can see them chat in
14:58
in Slack. So I think the
14:59
the
15:00
right now at least the communication is very
15:02
explicit and We can see it all
15:05
so I think if
15:06
I think if there's like a first class
15:07
way for the agents to communicate with each other
15:09
they're going to use that
15:11
rather than try to do something like back channeling
15:13
I mean, it feels like you have to get them
15:15
you have to really instruct
15:17
or post-train or
15:19
have like pretty aggressive instructions
15:21
to get them to do malicious things
15:23
actually
15:24
I mean
15:24
I've been using obliterated models lately
15:26
and they're super interesting and then that
15:29
you know, that's probably where
15:31
the thing is like the obliterated models coming
15:33
out of the box are not
15:35
you're, they're actually still very
15:36
very helpful
15:37
It's just that, if you instruct them
15:39
they won't refuse. And so at that point
15:41
like, that's where there's this black box around like
15:44
well, what context are you giving it?
15:46
And then
15:47
like that matters a lot because you could
15:49
actually give an obliterated model quite malicious instructions and because it
15:55
the, the, the obliteration is actually removing the refusal
15:59
like it, it will just do it actually
16:02
And so that's, that's the combo
16:03
I'd be worried about. But
16:05
you know, I mean
16:05
the reality is like it's very hard to
16:07
find
16:08
and set up an obliterated model and get it going
16:11
Yeah
16:12
I think people are
16:13
still really building trust with their bots
16:16
But I think for a lot of people
16:18
it's still quite early
16:19
And a lot of the way that they
16:20
build that trust is you start with something
16:21
really small. It's read-only
16:23
you know
16:23
has no real access to your tools
16:26
You see that
16:26
it did the work in the way you would expect
16:28
kind of take step two
16:30
And I think a lot of the step
16:31
one bots that have really taken off internally for us
16:34
Databot is one of them
16:35
I think the data science team is an
16:36
example of that. They constantly are getting Slack messages about
16:40
hey, I'm thinking about launching this feature
16:41
Can you do a quick analysis?
16:43
So they said, you know what?
16:44
This is taking a lot of our time
16:45
Can we just build a bot that does
16:46
this? And it's read access to our database
16:49
kind of can't take any action beyond that
16:50
and it can answer these questions
16:51
And it lives in Slack
16:52
So that was one
16:53
We had a similar case of kind of
16:55
vendor procurement
16:56
It's a lot of kind of
16:57
work there that probably
16:59
could be automated if someone like spent the
17:01
time to do it
17:02
And so, of course
17:02
that team
17:03
A lot of pull to build
17:04
a Grok bot that they added to Slack
17:06
and shared with the whole team
17:07
So that's kind of step one
17:08
I think step two is it can actually take action and
17:11
you know, it can
17:12
can really take on a task end to
17:13
end. Step three is maybe it
17:15
owns an entire
17:17
job responsibility or it kind of thinks about
17:19
an outcome constantly
17:21
I think we're still quite far away from those
17:22
but I think step one is
17:24
really necessary for these companies to build trust
17:26
I'm curious, what is the
17:28
for you, Lauren, what is your, your loop from
17:31
requests in Slack to PR?
17:33
I think when you did that presentation that
17:34
had over 4 million views on your dev process
17:38
you were pushing about 2,000 plus PRs per month
17:41
and now you're like over 4,000?
17:43
Yeah
17:44
Wow, definitely is wild. So
17:46
what do you do?
17:48
I, I can show some
17:49
like, examples
17:50
My loop is kind of interesting now
17:52
where I actually find myself talking to my Grok bots
17:56
mostly on Slack
17:58
and of course, X as well, when I'm helping, like, a customer or, you know
18:02
a Grok bot user
18:04
debug something or, you know, give them
18:06
some, ideas on how to solve a problem
18:09
Here on my screen
18:10
I have an example of me collaborating with my colleague
18:13
Samantha, on building out a feature together with my bot
18:17
Potato
18:18
and, you know, basically, I, I kind of started with saying
18:22
I gave it a link to some
18:24
like a video or something that I found on X
18:26
and I asked my bot
18:28
you know
18:28
what would it take for us to support
18:30
a feature like this?
18:31
And it went off and spawned a cursor
18:33
cloud agent
18:34
a cursor cloud agent
18:35
you know
18:35
has its own virtual machine as well as its own computer
18:38
so
18:39
it can actually run the Grokbot
18:42
app on its computer
18:43
and prototype and come back with videos and screenshots
18:47
And then, it actually did
18:49
So, this is quite powerful because now
18:53
you know, the conversation's happening in public
18:56
right? It's a, it's a public channel
18:58
it's a thread that anyone can
18:59
who's interested in this topic can join in
19:02
and give feedback
19:03
Process of iterating on a feature
19:06
together with someone is so much shortened by having these conversations
19:12
together, you know, in like a multiplayer setting versus
19:15
you know
19:15
if it was just me going off and building it
19:17
and I could certainly come up with something
19:19
but maybe the, the, the feedback process is
19:23
it comes later in the process
19:25
and the loop is sort of like
19:27
lengthened rather than shortened
19:29
And here I felt like with this particular example
19:32
the loop for us to collaborate was just
19:34
so much shorter. You know
19:35
we could see the screenshots
19:36
we could see the videos right away
19:38
How often would you jump back into cursor
19:40
and like sort of complete a PR?
19:42
Occasionally, I think. So not that often
19:44
I mean, you basically can stay in Crockbot
19:46
like stay up here
19:47
tell it what you need
19:49
you know, sort of give it direction
19:50
and like as an actual employee
19:52
it's like, it's just going
19:53
Yeah, I think for us
19:54
so the, the reason we can sort of do this
19:57
you know, workflow
19:59
Feature just like that
20:00
Yeah, we, we, we literally just like
20:02
I didn't even look at cursor for this project
20:04
We, I just, we just, we did it all in
20:06
in this Slack thread
20:07
It's got 200 messages. So it's quite a
20:09
it was quite a long
20:11
a lot of, of, of back and forth
20:13
But it, it, we came up with something that
20:14
you know, me, Samantha
20:16
and another PM were really like happy with
20:18
the result. We should ship this feature a lot faster than
20:21
you know, a longer process
20:22
And so the
20:23
the whole feature was built end to end
20:25
in that one Slack thread
20:26
with you and Samantha and Potato building it
20:30
Yeah. I mean, basically like this is
20:32
the gold standard for how to do it
20:34
And then I would say
20:36
maybe like five or 10% of engineers
20:38
or maybe 1%
20:40
of engineers period in the world are doing
20:42
this right now. So this is
20:44
literally something that should grow by
20:46
a hundred X. It's unbelievable that like
20:49
so few people are doing feature work this way
20:52
Yeah, I'll say like, you know
20:53
it took us actually quite a while to
20:54
get to that point
20:55
I feel like it's actually quite hard to
20:56
get to that point overnight
20:58
It involves You know
20:59
investment into making sure you have the right skills
21:02
you know
21:02
your code base has good guardrails and constraints so that
21:06
you know
21:06
the time that you would otherwise have to
21:08
spend on human code review
21:10
is encoded into your code base and your
21:13
you know
21:13
static analysis and your tooling so that the engineer's time
21:17
yeah, like PSAC
21:18
And so your engineer's time is freed up
21:20
to actually focus on the actual product
21:22
you know
21:22
the user experience
21:24
and the operational execution part of the job
21:27
You know
21:27
when you have so much more trust in
21:28
your agent's ability to write good code
21:31
because you've set up those systems
21:33
I've gotten to a point where
21:35
I essentially have my agents merge their own
21:37
pull requests. So in that Slack thread example
21:40
you know, like that was a
21:41
that was an example of
21:42
you know, me not reading the PR and this
21:45
I, we just looked at the
21:46
the artifacts that came out of it
21:47
So this cursor environment
21:49
does it already have PStack and is that
21:51
the way it sort of creates
21:54
you know, decides how to build it or?
21:56
Yeah. So, for internal
21:59
PSAC is installed into our repo
22:02
There's an internal version of PSAC
22:04
like instead of calling it potato mode
22:06
it, it's Loren mode
22:09
and it just started because
22:10
you know
22:10
I wanted a place to put my skills
22:12
And so, yeah, everyone
22:13
everyone in the company can actually use PSAC
22:16
and my cloud agents
22:17
that the bots spawn have the same access
22:20
to those skills. We haven't really explained what
22:21
PSAC is. Can you explain what
22:23
what PSAC is for people who don't know?
22:26
PSAC is actually inspired by GSAC
22:28
Oh, thank you so much
22:29
I'm honored
22:30
plugin
22:32
you know, I think
22:33
Gary, your, you know
22:34
GSAC was like quite inspiring in the sense of
22:36
I thought of it as like your personal set of
22:39
you know, chef knives. Oh, yeah
22:40
Right, like, you know, every chef when
22:42
when they get a new job
22:43
they go to a new kitchen
22:44
they bring their knives with them
22:46
And to me, you know, GSAC represented, you know, your opinionated take on how
22:52
you want your agents to
22:53
you know, help you
22:55
And so that inspired me of
22:56
to think about, you know
22:58
like what What a G stack look like for
23:00
you know, design more for like the engineering flows that
23:03
that I personally like
23:06
So I often tell people like you should
23:07
you should actually go and create your own
23:09
stack. Go fork it
23:10
you know, take our skills, combine them, you know
23:13
find what works for you
23:14
But I think the trust element is so
23:16
important. Like when you trust the
23:17
you know, your own tools
23:19
it
23:20
empowers you and your agents to do so
23:21
much more. I didn't really answer the question
23:24
but P stack stands for potato stack
23:25
I guess
23:27
And it's my set of skills that I've
23:28
open sourced on the cursor plugins repo
23:31
And it teaches
23:33
agents how to do
23:34
engineering work rigorously the same way that
23:37
I did when I was writing code
23:39
and it's kind of evolved a little bit and yeah
23:42
I guess it's gotten quite popular
23:44
Congrats. Yeah, it's awesome. Yeah, I
23:46
I am very inspired by it
23:48
And G stack is like pulling things from
23:50
it whenever I can. So
23:52
Yeah. Do you have a software factory that
23:54
like pulled in some ideas from P stack automatically?
23:57
Oh yeah, yeah
23:57
yeah. I mean now I use
23:59
Software factory is Cappy
24:00
and then
24:01
I have a captain that is sort of maps to
24:04
I think the Grok bot
24:05
the bot idea, but honestly
24:08
like now I think like maybe I need
24:09
to use a bot on top of Cappy
24:12
But yeah
24:12
I basically
24:13
can just
24:14
go to that captain and it's like the
24:16
master of my G stack
24:19
like sort of merge queue
24:20
like I
24:21
basically have like 10 or 15 threads working all the time
24:23
and then one of them is like Theo.gg
24:26
he really is T stack
24:28
and then P stack
24:29
and then obviously there's like lots of really
24:31
good inspiration out there
24:32
especially around design tools
24:35
And so it's really fun to just say
24:37
hey, go out, find what's out there
24:39
and like based on what you know about me
24:42
pull in the things that you think would
24:43
be most useful to improve my workflow
24:47
and then the really interesting thing about Cappy is
24:50
because all of my interactions with
24:53
my agent
24:54
my like coding agents are right there in the transcripts
24:57
it's developing and understanding. Like what I want
25:00
like when do I intervene
25:02
when do I say just do it
25:04
And so yeah
25:05
there's sort of this like very personal element
25:07
that's emerging
25:08
really through like the idea of memory
25:10
And so as you use the agents more
25:13
the more it knows what you want
25:15
Thank you for making P stack
25:16
And then both of us are
25:18
very
25:18
maligned and attacked on the internet by people
25:20
who refuse to believe that any of this
25:22
actually works. But you know
25:23
it's great to have a very brilliant compatriot
25:26
I'm not crazy. Like
25:27
this is actually something that is happening
25:29
And honestly
25:30
like I'm so convinced that in five or 10 years
25:33
or even like in two years
25:34
like it's going to be unfathomable that like
25:36
people are not doing it this way
25:38
But we're so early though still
25:41
One interesting
25:42
change that kind of
25:43
came up as you were sharing your screen
25:45
is it feels like in the past six months
25:47
so much of the work
25:48
engineering work
25:49
non-engineering work with agents is happening in Slack
25:53
And I think it has this interesting property
25:55
of since it's out in the open
25:57
in the past
25:58
the way that Lauren would work with coding
26:00
agents would have been entirely single player
26:02
private
26:03
The rest of the company would not have
26:04
benefited from
26:06
all of the crazy workflows that
26:08
they should totally learn from and adopt for their own
26:11
their own purposes. But now that it's happening in Slack
26:14
I've seen this, and maybe you have too
26:15
time and time again
26:16
where you're in some feedback thread
26:19
someone has a good idea
26:20
you spin off an agent
26:21
like at potato, you know, Lauren mode
26:24
And then the next day I see someone
26:25
who was in that thread do the same
26:26
thing
26:27
And I think
26:28
we're not just going to see that in
26:29
engineering. I think we're going to see that
26:30
in the way that people work with AI
26:32
in general. And the more that the stuff is in public
26:34
I think more
26:35
the more of the rest of the company
26:37
that can get lifted up by
26:39
the top 1% power users who are very tooling minded
26:42
very engineering minded
26:43
and know exactly how to use these tools
26:46
YC's next batch is now taking applications
26:49
Got a startup in you?
26:50
Apply at ycombinator.com
26:52
Apply
26:53
It's never too early
26:54
and filling out the app will level up your idea
26:57
Okay, back to the video
26:59
How much of your
27:01
of your, like
27:01
hours spent using Grokbot
27:03
is in the Grokbot app versus in Slack
27:06
or some other
27:07
service area to talk to the bots there?
27:09
I think a lot of the engineering work
27:11
that I do now
27:12
actually happens on Slack and X
27:15
So not all my bots have been added to Slack
27:17
so there are still a few that I kind of
27:19
that are more personal. Like I have a
27:21
I guess like a sort of personal assistant
27:23
bot that helps me with like
27:24
you know, picking my lunch
27:26
I don't need to add that to Slack
27:27
That's kind of, you know
27:28
personal or something that looks at my email
27:30
That's not a team bot
27:31
That's not a team bot
27:32
I guess we could
27:32
That's a personal bot. We could probably get
27:34
like a shared lunch bot
27:35
sure
27:36
I think that the nice thing as well
27:37
is like you can
27:37
you, you
27:38
you decide and choose what you want to
27:40
add to Slack
27:41
so my engineering bots tend to be
27:44
on Slack
27:45
so because I just find it so much
27:46
more ergonomic to Just at the bot in the Slack thread
27:51
you know
27:51
versus me having to go copy paste the link
27:54
So it
27:55
it just felt more
27:56
natural to bring them in that environment
27:58
I guess it really fits with your analogy
28:00
of bots as employees. Like if Slack is
28:02
where the work happens
28:03
that's where the human employees are
28:04
It's also where the like bot employees should
28:06
be so that they can work alongside us
28:08
Totally
28:08
And we don't shy away from that at
28:10
all. Like I think when we see people
28:11
use Grokbot in Slack
28:13
we love that. I think that's
28:15
that's where work happens
28:16
right?
28:16
And I think similarly
28:18
you know
28:18
in the future they should join meetings
28:19
They should be in your Google Docs
28:21
They
28:21
our goal is not to
28:23
rebuild all of B2B SaaS
28:25
you know, like other companies are
28:26
are doing a great job there
28:28
Our goal is to build the thing that
28:29
can access those tools and work inside of them
28:32
and just fit naturally into the way
28:34
a company already works
28:35
I guess one thing that
28:37
I'm curious about is
28:39
where do you think the future Let's say 2027
28:44
how should the best AI build engineers build?
28:47
So my hope for 2027
28:49
I think is like much more
28:50
you know, like we've
28:52
maybe accomplished more of this vision that we have for
28:54
for your Grok bots being a
28:56
like a coworker, you know, to have their own identity
28:59
their own telephone number
29:01
credit card, like basically almost anything
29:05
a human can do
29:06
I think we would want bots to do
29:07
And
29:09
I'd be very curious to see like how
29:11
this evolves
29:13
you know, I asked this question on my
29:14
on my X profile the other day as well
29:16
of like
29:17
what is the final form of agent UI?
29:20
And I, I posed the question of like
29:22
is it a chat interface or is it
29:24
something else?
29:25
one thing actually pretty interesting that I saw
29:27
yesterday on my X timeline
29:30
is a post from the Neuralink
29:32
team, and just came out with this new product called Telepathy
29:37
that I think, if I, if I'm not wrong
29:40
is along the lines of
29:41
in the future, an agent Being able to
29:44
essentially, like, read your brain, your mind through, like, you know
29:48
a brain computer interface
29:50
and I'm, I'm super interested in
29:52
like, what the future will bring when the
29:55
the sort of barrier between you
29:56
and how you chat and communicate in tent
29:58
with your bots becomes so much more fluid
30:01
That can happen, like
30:02
literally at the speed of thought
30:04
So that's personally, I think
30:06
could be
30:06
very interesting to see how that evolves
30:09
But I don't know
30:09
what about you, Roman?
30:11
We talk a lot about
30:12
the software factory and then
30:15
the factory for other parts of the business
30:17
And I think it's worth asking
30:18
like, what is different between those two?
30:21
And there are a lot of differences
30:22
And I think the next few years will
30:24
be on the product side
30:26
on the model side
30:27
like really trying to bridge the gap
30:29
I think one of the big differences is code
30:31
you know, it compiles or it doesn't
30:32
and it passes CI or it doesn't
30:34
And I think there's just more
30:36
infrastructure built around the way that code is written
30:38
tested
30:39
deployed
30:40
than knowledge work
30:41
and then like the rest of the company
30:42
There's just no real analog for those things
30:46
I think second is code
30:48
like we know how much
30:49
AI models benefit from the right context
30:52
And
30:53
code and software has the benefit of
30:55
the context living in a code base
30:57
But for the rest of the company
30:58
so much important information is stuck in people's
31:00
heads
31:01
It's stuck in pen and paper
31:03
It's stuck in some filing cabinet or some thing
31:06
you know, behind a password that you forgot
31:08
whatever
31:09
And I think we're seeing the limits of
31:10
that. That too
31:12
And so I'm pretty interested to see the way that especially
31:15
you know
31:15
a YC startup
31:16
that is starting very much from
31:18
first principles from scratch
31:20
and
31:20
can maybe build their company in a way
31:22
that it's not just a software factory
31:24
but is a factory across the whole company
31:26
Like what changes
31:28
will that have on company building?
31:29
What decisions will they make that
31:31
a company that was founded two years ago
31:33
maybe would not have
31:34
have made the same decision or trade-offs?
31:36
so I'm, I'm particularly interested to see that
31:38
We've been talking about
31:39
a bit about this concept of the AI native company
31:42
right?
31:43
And I think this
31:45
Got a lot more attention when Jack Dorsey
31:47
made that post
31:48
about how really companies should be
31:51
Block needs to be a
31:52
an AGI for, payments. Right
31:56
It's like you're
31:56
you're building a mini AGI and the whole
31:58
point of it is like Karen feeding of
32:00
the AGI that does this function for humanity
32:03
Right. I mean, the, the way
32:05
Jack Dorsey talked about it is like
32:07
you know, organization will get
32:09
get rid of
32:09
pretty much all middle management
32:11
and it would
32:12
basically boil down to just three roles
32:14
You would have sort of the
32:16
everyone becomes an IC because you just
32:17
ship everything. Everyone comes with outcomes
32:20
as opposed to the marketing team or growth
32:22
team working with engineering to build something
32:24
They could just build it themselves
32:26
And then the next role would be sort of the
32:28
the DRI that owns the outcome
32:30
It's almost like orchestrating all the ICs
32:32
And then you have this new role
32:34
which I think it might be just temporary for now
32:36
the player coach
32:37
which is someone that's super AI-pilled and teaches
32:39
everyone. Sounds like you are that role pretty much
32:41
Lauren. And then it's like
32:42
oh, show me the way
32:43
and then you level up the whole org
32:46
because it's like
32:46
this is how you call potato bot or matcha bot
32:50
right?
32:50
One of the cases was their customer success team using GrokBot
32:55
these team bots
32:56
So it's not just a
32:57
single player
32:58
thing that everybody on the team works with
33:00
It's really this colleague that is built and
33:02
inserted into Slack
33:04
And I think in this case
33:05
it fields every customer request
33:07
I think they get hundreds of messages from customers a minute
33:11
way too much for a human to parse through
33:14
and triages it
33:15
assigns it to the right human
33:17
I think for a lot of these tickets
33:19
AI could probably complete it end to end
33:21
And so for many of those
33:22
GrokBot does
33:23
and then it escalates it to the right
33:24
DRI
33:25
But I think your general point of
33:27
how companies can retool themselves for this moment
33:31
I think you're spot on that
33:33
I think a lot of companies are going
33:34
to spend more time on engineering objective functions of
33:40
you know, in a world where execution is kind of cheap
33:43
or you could spin up
33:44
a hundred coding agents or a hundred Grok
33:46
bots to like run at a problem
33:48
it really matters that you define the problem
33:50
well
33:51
And the
33:52
gap is probably not going to be an execution failure
33:55
it will probably be a
33:57
intense specification failure or a goal failure
34:01
And how you quantify these things
34:02
I think you can quantify a performance improvement of
34:05
you know, your desktop client
34:07
but I think it's a little bit harder
34:09
to maybe define
34:10
what an improvement to like
34:11
your marketing would be
34:12
And I think a lot of companies are
34:13
going to have to think about what those KPIs are
34:16
how you can point your agents in that direction
34:18
who's responsible for defining those things
34:20
And I think we've been living in
34:22
a bit of an execution bandwidth constrained world
34:25
and now we're going to shift to execution
34:27
abundance
34:28
and the types of questions that companies think
34:30
about will probably be quite different
34:32
Maybe that
34:32
Scarce resource is going to be more that
34:36
higher executive function of what to do and
34:38
why
34:39
And this is where
34:41
more
34:42
builders are going to have more the founder
34:44
traits of really knowing
34:47
translating user intent well into what
34:50
the bots need to do
34:51
Because I think that
34:52
that
34:52
there's a big gap between like the best
34:54
AI builders to
34:56
the so-so
34:56
ones is really how you translate that intent
34:59
super well. And
35:01
there's a lot of tips and tricks of what you did
35:02
Laura. I mean, your video is pretty impressive
35:04
Is there like one
35:06
One super important, just like if you do this
35:09
you'll get started. Like I think
35:11
was it like your
35:12
you know
35:12
what's the minimal form of a loop?
35:14
This is the smallest loop essentially
35:16
Like so agents already
35:17
are able to write code
35:18
Maybe it's not good. Let me
35:20
scratch it out for now
35:22
So like, you know, like naively if you're using agents
35:25
you can definitely write code
35:26
You can run the code and
35:28
you know, today that
35:29
that if you haven't spent a lot of
35:30
time in your agent setup
35:32
you may be the person
35:33
you know, running the code
35:34
And so like this is an actually a loop
35:37
really because, right, right now this is like the
35:40
the missing piece. The agent can't do this
35:43
but you were doing it
35:44
So the, the smallest loop I find is
35:46
you know, like, actually focus on this part
35:49
How do you teach your agent to
35:51
you know, run the code empirically
35:53
you know, take CPU traces, you know, like performance metrics
35:57
like run the code
35:59
and run the application like a real user
36:00
would
36:01
The reason why I think it provides a
36:02
loop is because the
36:03
agent now has the ability to check its
36:05
own work
36:06
So it can find the issues with the
36:07
code that it wrote
36:08
then it can write better code or more correct code
36:11
and then it can keep
36:12
looping
36:13
until it produces a result that you want
36:16
I took out the good part for a reason
36:17
because I think that
36:19
you know, when you have a loop like this
36:20
it doesn't guarantee that the quality of the code is
36:24
is good, but it certainly will be correct
36:27
You know
36:27
correct in the sense that it does the
36:29
thing that you want it to
36:30
but maybe it's like very slow
36:32
or very resource inefficient
36:34
And that's where, you know
36:35
the engineers sort of come into play
36:37
It's like
36:37
the new job of a software engineer is
36:39
to figure out how to get This to happen
36:42
Right
36:42
but how do you get agents to write
36:44
good code by default?
36:46
and so the loop gets smarter and smarter
36:49
and, when you're happy with it
36:52
I think that's the point where you can start thinking about
36:54
you know, the other parts of the loop
36:56
like, you know, the outer loop part
36:58
where, you know, like, once this part is really good
37:01
you can now do things like
37:04
for example, just a, a simple example here
37:06
you have an alert for some metric or monitor that's
37:10
you know, going below a certain threshold
37:13
Now an agent, like a graph bot
37:15
a cursor automation
37:17
could be fired off as a result of that alert
37:20
and then it can go off and use the same
37:22
you know
37:23
tools and skills that you've provided for the basic loop
37:26
and it can repro
37:27
you know, reproduce the issue in its own machine
37:30
and then, you know, like, it runs the code
37:31
and then it finds the issues of the code
37:33
and it writes the code
37:34
and then now you have this
37:35
like, self, sustaining loop, and that frees up my time
37:39
right?
37:40
From, you know
37:41
being babysitting every agent to now orchestrating multiple agents
37:46
and I even now have my bots
37:49
like I've gone another level where I
37:51
you know, like I said
37:51
we have this new feature called projects
37:53
that lets you have a project

Watch these videos on your TV

XTTV brings video content from X to your living room. Download now and start watching on the big screen.