WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=aeTb5BdmTTc

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/aeTb5BdmTTc.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.400 --> 00:00:15.920
Today, I'm going to be talking about agents code bases and teams.

s3
00:00:15.920 --> 00:00:20.040
Essentially, how do you get your team to actually ship together with agents?

s4
00:00:20.040 --> 00:00:31.000
And like, I think for the longest time, like the one thing that's bugged me is like there's so much content about, you know, how do you set up your own code base to like work well with agents, you know, like what skills do you add?

s5
00:00:31.000 --> 00:00:33.520
You know, this skill's better, that setup's better.

s6
00:00:33.520 --> 00:00:39.600
But it all seems to break the moment you like actually try to use it with your team in your actual production setup.

s7
00:00:39.600 --> 00:00:42.320
And like for individual repos, it kind of makes sense.

s8
00:00:42.320 --> 00:00:47.840
But the moment you actually try to use it with your own like team setup, it tends to break.

s9
00:00:47.840 --> 00:00:53.160
And like I think over the past few months, I've figured out how to make it work with a team of folks.

s10
00:00:53.160 --> 00:01:00.080
So like I was leading a team of like 10 people over the last few months, and I think we found like a good solution, and I want to share that with you guys.

s11
00:01:00.080 --> 00:01:04.760
But before we get into that, I just want to recap like, you know, what's been the journey that we've been on.

s12
00:01:04.760 --> 00:01:05.239
Right?

s13
00:01:05.239 --> 00:01:11.920
So we started, you know, code coding agents took off, and a few people got really, really good leverage, right?

s14
00:01:11.920 --> 00:01:15.000
And I think all of us were kind of asking like, is this AGI?

s15
00:01:15.000 --> 00:01:17.880
You know, like did we achieve it?

s16
00:01:17.880 --> 00:01:24.720
Then, you know, companies took that and said, "Well, if one person can do so well, let's just get everyone and let's mandate it."

s17
00:01:24.720 --> 00:01:27.000
And like token max, right?

s18
00:01:27.000 --> 00:01:30.760
And that was clearly a a galaxy brain moment, right?

s19
00:01:30.760 --> 00:01:36.040
And then the inevitable happened, you know, AI slop shipped, and there's a bunch of sev twos.

s20
00:01:36.040 --> 00:01:40.560
I'm not going to name which companies, but you know, essentially you saw people kind of retracting.

s21
00:01:40.560 --> 00:01:45.040
They said, "I don't think I don't think this is the best option here."

s22
00:01:45.040 --> 00:01:48.160
And eventually model prices you know, climbed.

s23
00:01:48.160 --> 00:01:52.080
We saw, you know, people figured out that tokens have to be paid for.

s24
00:01:52.080 --> 00:01:54.960
Like you just can't token max your way through life.

s25
00:01:54.960 --> 00:01:56.800
And budgets got bolted on.

s26
00:01:56.800 --> 00:02:01.640
And you know, essentially money is being lit on fire, and the money has to come from somewhere.

s27
00:02:01.640 --> 00:02:07.160
So, given this journey, I want to actually like this is the enterprise journey, right?

s28
00:02:07.160 --> 00:02:09.880
And what does that do for a single developer?

s29
00:02:09.880 --> 00:02:15.640
And I think this is an important framing cuz it really talks about people as a part of a team, right?

s30
00:02:15.640 --> 00:02:17.560
And I want to look at it from like two axes.

s31
00:02:17.560 --> 00:02:22.440
So, there is the fear axis where people lie on the spectrum, right?

s32
00:02:22.440 --> 00:02:24.360
They're coming It's like Is it coming from my job?

s33
00:02:24.360 --> 00:02:25.840
Like am I going to be out of a job?

s34
00:02:25.840 --> 00:02:27.760
Or is it like a really handy tool?

s35
00:02:27.760 --> 00:02:29.760
And they're really, you know, they're not that fearful.

s36
00:02:29.760 --> 00:02:32.760
Versus like the confidence they have in how much they're executing it.

s37
00:02:32.760 --> 00:02:39.640
So, they can either use it a lot or they can use it, you know, not that much because they don't really know how to use it that well.

s38
00:02:39.640 --> 00:02:43.560
Now, when we started, like people said, "Oh, you know, what is this?

s39
00:02:43.560 --> 00:02:44.280
Is this the end?

s40
00:02:44.280 --> 00:02:45.080
Like am I needed?"

s41
00:02:45.080 --> 00:02:46.840
And fear was pretty high.

s42
00:02:46.840 --> 00:02:50.200
Utilization was pretty low cuz people didn't really know how to use it.

s43
00:02:50.200 --> 00:03:01.959
And then when a few people got outsized leverage, you saw, you know, early adopters basically people saw them and people said, "Okay, well, it looks like I'm still kind of needed if I figure out how to use this thing.

s44
00:03:01.959 --> 00:03:04.640
So, let me actually try using it, right?"

s45
00:03:04.640 --> 00:03:10.120
And then we saw, you know, mandates and token maxing, and people kind of got a little skeptical.

s46
00:03:10.120 --> 00:03:14.480
Like confidence stayed the same, but people tried to use it a lot more, right?

s47
00:03:14.480 --> 00:03:23.880
And then we realized, uh you know, there's a bunch of slop shipping, there's self-tours, and I, you know, it's like I'm not really that scared cuz, you know, it just slip it just ships slop.

s48
00:03:23.880 --> 00:03:25.440
Like I'm going to still be needed.

s49
00:03:25.440 --> 00:03:29.640
And they don't even know how to use it that well cuz like now the confidence is greater, right?

s50
00:03:29.640 --> 00:03:39.239
And so you got to figure out how to get people from all the way wherever they are on the spectrum to where they're not fearful and they're actually using it a whole lot more.

s51
00:03:39.239 --> 00:03:48.959
And this is the framing that I want to like I want everyone to keep in mind as they're actually trying to get a team to adopt good like AI usage and good AI patterns, right?

s52
00:03:48.959 --> 00:03:51.880
And so, the The is, you know, what does it take?

s53
00:03:51.880 --> 00:03:54.400
Step one, create a Cloud MD.

s54
00:03:54.400 --> 00:03:57.400
Step two, add some skills.

s55
00:03:57.400 --> 00:03:58.600
Is that it?

s56
00:03:58.600 --> 00:04:00.600
Do we like solve it?

s57
00:04:00.600 --> 00:04:08.280
You know, I think we all know like you guys are here because like clearly life's not that simple and like stuff's messy, right?

s58
00:04:08.280 --> 00:04:11.280
And I think a few people might ask, you know, why doesn't this work?

s59
00:04:11.280 --> 00:04:13.920
Like isn't that what like everyone does?

s60
00:04:13.920 --> 00:04:22.040
And I want to like just talk about a few things you might see that actually indicate that yeah, like this isn't working.

s61
00:04:22.040 --> 00:04:27.440
So, the first thing is if you're babysitting your agents, it's not the right setup, right?

s62
00:04:27.440 --> 00:04:29.040
And you got to realize that.

s63
00:04:29.040 --> 00:04:34.080
If you're seeing people in your team babysitting their agents, something's wrong.

s64
00:04:34.200 --> 00:04:41.280
Is, you know, one of the things that I heard a lot was is, you know, insert whatever latest model there is being really dumb today.

s65
00:04:41.280 --> 00:04:42.800
The model didn't change, right?

s66
00:04:42.800 --> 00:04:44.919
The hardness may have changed underneath.

s67
00:04:44.919 --> 00:04:53.880
But if it's really like that's that's acceptable to like small changes in the hardness, clearly your own code base isn't set up well.

s68
00:04:54.360 --> 00:04:56.240
It's silently burning context and money.

s69
00:04:56.240 --> 00:05:06.080
Like you don't realize it, you know, you go you blow through like 500k context, you might go to like 750k million and hit auto compact even though you're not doing like a really complicated task.

s70
00:05:06.080 --> 00:05:08.080
Clearly something's wrong.

s71
00:05:08.080 --> 00:05:14.800
You know, if you have long ass sessions needing constant intervention, there's still something that's wrong.

s72
00:05:15.000 --> 00:05:20.520
If you're getting a constant slot factory, obviously know like things are things are not good.

s73
00:05:20.520 --> 00:05:25.240
And if you're asking if you find yourself asking, you know, how are these other companies shipping so fast?

s74
00:05:25.240 --> 00:05:30.320
Like how are model companies releasing models at like a month and a half two-month cadence?

s75
00:05:30.320 --> 00:05:34.880
Clearly they they have something which we don't, right?

s76
00:05:34.880 --> 00:05:40.120
And so, I guess everyone's kind of thinking so, you know, how do we solve this correctly?

s77
00:05:40.120 --> 00:05:45.120
And so, I think the first thing to realize is we need to frame it correctly, right?

s78
00:05:45.120 --> 00:05:47.400
It isn't really an IC's job.

s79
00:05:47.400 --> 00:05:48.640
It's a job for leadership.

s80
00:05:48.640 --> 00:05:51.080
It's a job for the company, right?

s81
00:05:51.080 --> 00:05:58.520
Making engineers work well with their agents is truly the most impactful thing you could do as an organization.

s82
00:05:58.520 --> 00:06:06.320
Because that's going to enable your engineers to ship faster and with confidence and avoid, you know, a lot of incidents.

s83
00:06:06.320 --> 00:06:15.280
If we live in this figure it out for yourself paradigm, like people are going to get outsized productivity, some people aren't, and like the people who are generating like 10 PRs a day

s84
00:06:15.280 --> 00:06:20.280
are going to like look like, you know, gods compared to people who are shipping like one to two.

s85
00:06:20.280 --> 00:06:24.280
And the one to two PR people are actually going to get left with the review burden.

s86
00:06:24.280 --> 00:06:26.200
And that's actually a really, really bad thing.

s87
00:06:26.200 --> 00:06:36.600
Cuz now not only can they not ship, they're going to actually see bad code and then curse the agents and hence not be able to get on to the let's ship 10 PRs, right?

s88
00:06:36.600 --> 00:06:39.160
And so it's really important to do this.

s89
00:06:39.160 --> 00:06:43.240
If it's a problem facing the team, there's a few things you can do, right?

s90
00:06:43.240 --> 00:06:49.680
The most impactful things that you can do to set up your code base to like make it work well require team buy-in.

s91
00:06:49.680 --> 00:06:55.480
You can't just like if you want to change the way your code base is organized, you can't do that as an IC, right?

s92
00:06:55.480 --> 00:07:01.480
And if you're treated as a leadership problem, then you can do things like this.

s93
00:07:01.480 --> 00:07:05.440
So, the other thing this needs is harness engineering, right?

s94
00:07:05.440 --> 00:07:06.360
Per code base.

s95
00:07:06.360 --> 00:07:13.760
And I think there's a lot of content on this, so I just want to talk about a few principles, but I don't want to like make this talk about that because there's a lot of smart people.

s96
00:07:13.760 --> 00:07:15.000
You're at AI engineer, right?

s97
00:07:15.000 --> 00:07:21.400
Like this this conference is all about people telling you how to you best set up, you know, your code base to make things function well.

s98
00:07:21.400 --> 00:07:25.640
So, I don't want to talk too much about this, but there's a few key principles here.

s99
00:07:25.640 --> 00:07:27.480
Smart prompt injection is one of them.

s100
00:07:27.480 --> 00:07:36.760
You want to treat your entire code base as one way to that so that you're able to like smartly prompt inject the model with just the right context at just the right time.

s101
00:07:36.760 --> 00:07:38.800
Without you needing to do it.

s102
00:07:38.800 --> 00:07:39.960
And that's the framing.

s103
00:07:39.960 --> 00:07:43.200
You want to be able to say, "Okay, I've set it off on this task.

s104
00:07:43.200 --> 00:07:48.160
It has like a map of how to find the things it needs at the time it needs it."

s105
00:07:48.160 --> 00:07:54.280
If it's looking at some code and that code has, let's say some documentation, it the documentation needs to live in the comments.

s106
00:07:54.280 --> 00:08:01.400
So, if it ever grabbed into that code, it reads the comment, goes to that file, finds all the information about it.

s107
00:08:01.400 --> 00:08:02.800
That's just one example.

s108
00:08:02.800 --> 00:08:04.880
The second is close the loop, right?

s109
00:08:04.880 --> 00:08:08.040
You got to make a self-healing system cuz slop is inevitable.

s110
00:08:08.040 --> 00:08:11.640
There is going to be some slop that's going to seep in.

s111
00:08:11.640 --> 00:08:19.320
But, you need to have a pipeline and a way to close the loop to remove the slop, to detect it, and to be able to like self-heal the system.

s112
00:08:19.320 --> 00:08:21.240
And then you need to iterate continuously.

s113
00:08:21.240 --> 00:08:22.560
And I can't emphasize this enough.

s114
00:08:22.560 --> 00:08:26.200
Like, you can't assume that you do this for a month and you're done.

s115
00:08:26.200 --> 00:08:28.800
Like, things are going to change constantly underneath.

s116
00:08:28.800 --> 00:08:32.960
So, you need to keep this as one of the things that you have to do as an organization.

s117
00:08:32.960 --> 00:08:36.360
And the third most important thing is treat it like a human problem, guys.

s118
00:08:36.360 --> 00:08:39.360
Like, this isn't It's not, you know, oh, it's this tool.

s119
00:08:39.360 --> 00:08:40.360
Like, people will figure it out.

s120
00:08:40.360 --> 00:08:41.719
Let's just mandate our way through life.

s121
00:08:41.719 --> 00:08:44.000
Like, that's just not going to work.

s122
00:08:44.000 --> 00:08:45.720
So, treat it like a human problem.

s123
00:08:45.720 --> 00:08:47.040
Fear is real.

s124
00:08:47.040 --> 00:08:48.400
Human emotions are real.

s125
00:08:48.400 --> 00:08:50.640
We should recognize it.

s126
00:08:50.640 --> 00:08:58.520
So, you know, enough gyan or as like it's more like the Hindi way to say like, you know, I'm enough prof like I'm kind of giving you sermons.

s127
00:08:58.520 --> 00:09:00.480
But, like, how do you really do this, right?

s128
00:09:00.480 --> 00:09:02.160
Like, these are all principles.

s129
00:09:02.160 --> 00:09:03.600
What's the real playbook?

s130
00:09:03.600 --> 00:09:14.400
So, here's what we did and here's like I'm not going to over like emphasize on this is the exact way to do it, but this is like roughly how we did it and you can take from it what you choose.

s131
00:09:14.400 --> 00:09:16.200
The first thing is do the basics, right?

s132
00:09:16.200 --> 00:09:17.600
You got to do them right.

s133
00:09:17.600 --> 00:09:22.480
Progressive disclosure, I can't emphasize this enough, is really, really powerful, right?

s134
00:09:22.480 --> 00:09:27.680
Find your best ICs and find what how they are making the code base work for them.

s135
00:09:27.680 --> 00:09:31.040
Take those practices and pass them or guide.

s136
00:09:31.040 --> 00:09:32.760
People can't live in their own practices.

s137
00:09:32.760 --> 00:09:34.280
And this is really hard for engineers to do it.

s138
00:09:34.280 --> 00:09:37.520
It's basically accepting that my setup is imperfect.

s139
00:09:37.520 --> 00:09:39.840
And like engineers don't like to hear that.

s140
00:09:39.840 --> 00:09:44.480
But you got to figure out a way to find those best practices and ship them across.

s141
00:09:44.480 --> 00:09:47.400
Make sure that that's kind of a shared setup.

s142
00:09:47.400 --> 00:09:51.520
The second thing we did was there's one high value skill that we invested in.

s143
00:09:51.520 --> 00:09:53.800
In our case it was this thing called ship it.

s144
00:09:53.800 --> 00:10:00.920
What it did was the moment you're done with your code, it takes care of everything from code done to PR ready for review.

s145
00:10:00.920 --> 00:10:10.600
Which means you got to open a PR, figure out your opinions, handle the handle all the comments, you know, handle all the PR descriptions, the merge comments, everything, right?

s146
00:10:10.600 --> 00:10:11.760
It handles CI failures.

s147
00:10:11.760 --> 00:10:14.280
It basically runs through these loops.

s148
00:10:14.280 --> 00:10:17.800
And what this meant was often the skill was running for over an hour.

s149
00:10:17.800 --> 00:10:23.960
And that scared people, but once you actually figure out once they saw the value, they get invested, right?

s150
00:10:23.960 --> 00:10:28.320
Because it's one skill which tells them, "Okay, this AI thing can actually work for me.

s151
00:10:28.320 --> 00:10:30.080
I don't need to constantly baby sit it.

s152
00:10:30.080 --> 00:10:32.320
I can trust it."

s153
00:10:32.320 --> 00:10:34.960
The third thing and really important is to close the loop, right?

s154
00:10:34.960 --> 00:10:36.920
So we wired issues and boards into the repo.

s155
00:10:36.920 --> 00:10:38.120
Like we added CICD.

s156
00:10:38.120 --> 00:10:39.520
We added gigantic reviews.

s157
00:10:39.520 --> 00:10:48.360
We have like a code gardener that actually goes back and looks through a whole bunch of like every night it'll run and look at the code and check if something like not organized correctly.

s158
00:10:48.360 --> 00:10:52.800
What does correct organization mean will depend on your code base.

s159
00:10:53.760 --> 00:10:56.240
Get people invested and I can't emphasize this enough.

s160
00:10:56.240 --> 00:10:58.240
You have to win over the skeptics.

s161
00:10:58.240 --> 00:11:01.640
It's really easy to say like the skeptic is just someone who's scared.

s162
00:11:01.640 --> 00:11:07.720
It's really hard to get them to buy in, but if you can get them to buy in, you know you're doing something right.

s163
00:11:07.720 --> 00:11:15.560
You have to get them to be able to edit and play with the shared setup cuz that's the true way you know that they're actually invested, right?

s164
00:11:15.560 --> 00:11:19.400
And this is where you got to it ensure you're iterating constantly.

s165
00:11:19.400 --> 00:11:27.280
If people are like and this is the hardest thing for engineers again cuz you're basically saying I'm never going to like get to perfection in my setup.

s166
00:11:27.280 --> 00:11:29.000
But you got to be okay with that.

s167
00:11:29.000 --> 00:11:37.960
You have to do it and you You to treat it like, you know, X% of your IC time is probably going to be spent on it trading on this thing, which is not going to lead to like meaningful PRs

s168
00:11:37.960 --> 00:11:42.640
like up front, but it's useful and it's worth it.

s169
00:11:42.640 --> 00:11:44.760
And, you know, I don't want to say this is perfect, right?

s170
00:11:44.760 --> 00:11:47.160
Like we faced a ton of issues while doing this.

s171
00:11:47.160 --> 00:11:49.360
And I'm just going to walk you through some of them.

s172
00:11:49.360 --> 00:11:52.920
But it's an iteration loop, so you got to treat it like a piece of feedback.

s173
00:11:52.920 --> 00:11:54.920
So, what are the problems we hit, right?

s174
00:11:54.920 --> 00:11:55.800
There are too many issues.

s175
00:11:55.800 --> 00:12:05.520
Like when we started like we blew up to like 4 or 500 issues uh without like, you know, just I think within like a couple of weeks, which is a crazy number for like a repo.

s176
00:12:05.520 --> 00:12:11.280
And then, you know, there are so many like different agents all trying to create issues cuz they've not been wired correctly.

s177
00:12:11.280 --> 00:12:12.160
There's a lack of agreement.

s178
00:12:12.160 --> 00:12:17.440
People, you know, as soon as people saw like, "Oh, this isn't working perfectly or the way I expected it."

s179
00:12:17.440 --> 00:12:19.080
There is super easy for them to say, "You know what?

s180
00:12:19.080 --> 00:12:21.320
I'm just going to go back to babysitting my agent."

s181
00:12:21.320 --> 00:12:22.480
You don't want that.

s182
00:12:22.480 --> 00:12:28.200
You want to actually take their feedback and put it back into the skill and improve the skill.

s183
00:12:28.600 --> 00:12:29.640
Agents are taking too long.

s184
00:12:29.640 --> 00:12:32.280
This is actually one of those expectation setting things.

s185
00:12:32.280 --> 00:12:33.720
It's good if agents take too long.

s186
00:12:33.720 --> 00:12:38.880
That means you can actually go off and do other things and you have confidence that they're doing the right thing.

s187
00:12:38.880 --> 00:12:44.839
At the end of the day like the moment we hit this reasoning paradigm, the longer the agent like thought, the better its output.

s188
00:12:44.839 --> 00:12:51.240
You can treat like it's It's almost like a similar mindset for your entire code base and for your skills.

s189
00:12:51.680 --> 00:12:54.600
There's going to be merge hell and like we just have to deal with it.

s190
00:12:54.600 --> 00:12:57.560
We have to figure out a way to deal with this.

s191
00:12:57.560 --> 00:13:00.400
There is going to be slop when you're going to write experiments.

s192
00:13:00.400 --> 00:13:02.680
Treat it like it's own thing, right?

s193
00:13:02.680 --> 00:13:06.120
What we said was okay, people are generating this code, but it's not relevant.

s194
00:13:06.120 --> 00:13:07.240
It's not going to be shipped.

s195
00:13:07.240 --> 00:13:09.200
It's just an It's a prototype.

s196
00:13:09.200 --> 00:13:10.200
So, treat it like one.

s197
00:13:10.200 --> 00:13:15.480
Get it to opt out of all the rigorous other standards you've got across your code base.

s198
00:13:15.480 --> 00:13:18.040
And realize people vary on the spectrum, right?

s199
00:13:18.040 --> 00:13:22.880
And depending on the day, depending on like what they're going through, they're going to vary on the spectrum.

s200
00:13:22.880 --> 00:13:27.080
You have to be able to talk to them and figure out, "Hey, okay, why why are you facing this?

s201
00:13:27.080 --> 00:13:31.720
Like is it if the model changed, the hardness changed, again you need to go revisit something.

s202
00:13:31.720 --> 00:13:33.960
Figure that out.

s203
00:13:33.960 --> 00:13:46.000
And I think the biggest like the easiest way to say this is instead of saying like the model is so dumb, like we have to ask how can I make it smarter or how can I edit like and not This is where I have like crossed out the my.

s204
00:13:46.000 --> 00:13:47.440
It's not a personal setup.

s205
00:13:47.440 --> 00:13:51.320
It's the shared setup that you have to invest in.

s206
00:13:51.440 --> 00:13:53.040
And it's a mindset, right?

s207
00:13:53.040 --> 00:13:55.320
You have to go full send.

s208
00:13:55.320 --> 00:13:57.280
And like I want to end you with I want to end with this.

s209
00:13:57.280 --> 00:14:05.000
Like, I learned skiing like a couple years back and the first like the hardest thing for me was you actually have to commit to it.

s210
00:14:05.000 --> 00:14:07.040
If you're pizza braking, you're going to crash.

s211
00:14:07.040 --> 00:14:08.040
Like, no matter what.

s212
00:14:08.040 --> 00:14:15.760
You have to commit to the speed in order to actually get and feel like, okay, that's how I can turn and that's how I can truly ski.

s213
00:14:15.760 --> 00:14:17.520
And so, I'm going to leave you with this.

s214
00:14:17.520 --> 00:14:21.480
Like, just be okay with be okay with failing.

s215
00:14:21.480 --> 00:14:24.440
You have to, you know, go full send and be okay with falling.

s216
00:14:24.440 --> 00:14:25.960
It's fine.

s217
00:14:25.960 --> 00:14:31.920
The point is to be able to recover from that and that will allow you to truly feel the AGI.

s218
00:14:32.320 --> 00:14:36.800
Yeah, well, that's that's me and I'm happy to take any questions.

s219
00:14:40.839 --> 00:14:42.839
Yep.

s220
00:14:45.720 --> 00:14:47.880
Uh so, the I'm going to repeat the question for the recording.

s221
00:14:47.880 --> 00:14:51.000
Uh strategies that you found best for progressive disclosure.

s222
00:14:51.000 --> 00:14:52.920
So, I think a couple things, right?

s223
00:14:52.920 --> 00:14:58.400
Uh the first thing is even in your skill.md files, don't overload it.

s224
00:14:58.400 --> 00:15:03.960
Like, we've kind of set a hard limit for like 100 lines in your skill.md cuz your skill is really a folder.

s225
00:15:03.960 --> 00:15:05.560
So, that's step one.

s226
00:15:05.560 --> 00:15:16.440
Make sure and I think I spoke about this during the talk, but when you have like some code that requires like a runbook, make sure the runbook is reflected in the comments

s227
00:15:16.440 --> 00:15:27.400
so that if a code if somehow the code like the agent figures its way into like grapping into the code base and find that file, it knows I need to go look at this for like all the description of how this is relevant.

s228
00:15:27.400 --> 00:15:28.080
Right?

s229
00:15:28.080 --> 00:15:36.640
You have to organize and I think this is why I talk about harness engineering because your entire code base can be set up to encourage progressive disclosure.

s230
00:15:36.640 --> 00:15:40.960
Don't overload your cloud MD or your agent's MD file into like one big thing.

s231
00:15:40.960 --> 00:15:50.680
You want to make sure that like it's a thin index that can point through the right files and that's what the agent gets in its like first prompt cuz that's what gets loaded when it starts to work.

s232
00:15:50.680 --> 00:15:56.720
So these are like some really powerful strategies that and the way you know this is working is when you give it a prompt.

s233
00:15:56.720 --> 00:15:59.160
When you give it the first prompt, see what it's doing.

s234
00:15:59.160 --> 00:16:02.120
Is it grabbing or does it know where to go?

s235
00:16:02.120 --> 00:16:04.680
But how much context is it burning when it like immediately?

s236
00:16:04.680 --> 00:16:10.280
So is it like I think like 20, 25K tokens get taken anyway, but like how much more is getting added?

s237
00:16:10.280 --> 00:16:13.360
If you're coming to like 40K, 50K, like something's wrong.

s238
00:16:13.360 --> 00:16:15.840
That's not really progressive disclosure.

s239
00:16:15.840 --> 00:16:21.480
So you have to figure out these boundaries and then like based on this it's an iteration cycle.

s240
00:16:31.600 --> 00:16:31.800
All right.

s241
00:16:31.800 --> 00:16:34.160
Well, if there there aren't any other questions, feel free to find me.

s242
00:16:34.160 --> 00:16:40.920
Happy to talk about harness engineering in general or like anything else, but yeah, thank you for listening.

s243
00:16:54.606 --> 00:16:56.606
[music]