AI Engineer World's Fair 2026

The Loop Is the Product — Roland Gavrilescu, Introspection

Read the talk

The Loop Is the Product

Roland Gavrilescu’s blueprint for improving agents connects execution, human judgment and production experiments: capture what a system learns in portable recipes, then optimize for work users value at a cost that makes sense.

From a talk by Roland Gavrilescu

At a glance

Ideas worth remembering

  • Build the feedback path as carefully as the worker: signals guide execution, verifiers check success, and artifacts inform the next improvement loop.

  • Preserve operating knowledge in versioned agent recipes: failures become evaluations, repeated behavior becomes skills, and human judgment remains part of the system.

  • Let humans calibrate the desired judgment and agents implement the evaluations. Then test whether users value that judgment before promoting a recipe change.

  • Establish that the output is valuable, then make its economics viable. Efficient execution only becomes a useful product advantage when customers want the work.

A car negotiation gives the loop an outcome

An agent that negotiates a car purchase has a concrete job: find available cars, compare prices, ask dealers for better offers and recognize when to close. Roland Gavrilescu, co-founder of Introspection, uses this example to introduce a product question that grew out of his work on agent infrastructure at xAI: how do you turn agents running long tasks into services that improve with their customers?

Source frame: A car negotiation gives the loop an outcome
Source frame: A car negotiation gives the loop an outcome

The discussion around agents had moved from improving models to improving harnesses—the surrounding software that lets a model use tools—and then to building loops. “The loop is the product” makes that progression concrete. The product needs a process that can act, receive useful information and decide what to do next.

In Gavrilescu’s account, an OpenClaw user assembled that process around car buying. Reddit discussions supplied price information; inventory searches established what was available; conversations with dealers produced competing offers. Putting dealers head-to-head gave the agent another action to take when an offer was unsatisfactory. A verifiable way to recognize an acceptable price supplied the stopping condition, after which the buyer could lock in the purchase. Gavrilescu reports that it worked, without giving a purchase price or savings figure.

0:210:27
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

Signals guide the work; verifiers judge the result

The next reference is the OODA loop: a way to organize observation and action in a changing environment. Gavrilescu connects it to agents calling tools and taking observations. Each action changes what the system knows, and the next decision depends on that new information. In the car example, a dealer’s response becomes input to the next negotiation.

Source frame: Signals guide the work; verifiers judge the result
Source frame: Signals guide the work; verifiers judge the result

Two parts of the loop deserve separate attention:

  • Signals: The information that guides the agent’s next action. Gavrilescu identifies signal quality as a determinant of the loop’s success rate.
  • Verifiers: The checks that establish whether an apparent success is actually correct. A result can look promising while failing the criteria that make it useful.

What changes when the result of one loop becomes input to another? The diagram shows the extra feedback path. The first loop does the work and checks its result. Its artifacts then give a second loop material for understanding what went well, what failed and what should change. Continuous improvement depends on making that return path useful.

How it fits togetherExecution feeds an improvement loop

Information that guides the worker

The verifier checks the work. The improvement loop uses the resulting artifacts to change the signals driving future work.

3:033:05
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:03 · section reference included

Store what the system learns as an agent recipe

“System distillation is the moat” names the second idea: preserve the knowledge accumulated while operating an agent. A run can teach you about its harness, profiles, evaluations, model, resources, tools or environment. If those lessons remain scattered across individual runs, they are hard to reproduce or improve deliberately. An agent recipe gathers them into something portable, versioned and able to evolve.

Source frame: Store what the system learns as an agent recipe
Source frame: Store what the system learns as an agent recipe

The analogy is to data recipes in reinforcement learning. As unwanted behaviors such as hallucination and reward hacking appear, researchers change the recipe that produces the system. Gavrilescu proposes applying that accumulated know-how to the broader agent system, including adjustments and human judgments discovered only after the agent begins acting in its environment.

Different observations become different parts of the recipe:

  • Failure patterns become judges and evaluations. Recurring mistakes supply criteria for checking subsequent runs.
  • Repeated behavior becomes skills and prompts. Useful ways of working become reusable instructions.
  • User frustration informs harness extensions and memory. Problems encountered during use change the support the agent receives.

The proposed recipe belongs to the company using it and remains independent of a particular model or provider. Putting it in a Git repository makes its evolution part of the product’s ongoing development. Here, distillation includes evaluations, skills, prompts and operating choices; the accumulated system knowledge has value even before any of it moves into model weights.

4:244:26
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:40 · section reference included

Pi recipes preserve the choices behind the agent

Introspection’s implementation uses the Pi harness, Harbor for evaluations and Git repositories for versioning. Gavrilescu describes an early release called Pi recipes, intended to be owned by the user and managed by agents. Version history gives those agents a way to track what changed and why. The release is presented as an early implementation of the recipe idea, rather than a demonstrated guarantee of reproducibility across every provider.

Source frame: Pi recipes preserve the choices behind the agent
Source frame: Pi recipes preserve the choices behind the agent

The knowledge worth carrying includes the maker’s taste: the judgments that distinguish a desirable result from a merely plausible one. Agents should calibrate themselves to those judgments. Reusing someone else’s recipe would therefore mean bringing along the choices that produced the system, including the reasons behind them.

That expands the recipe beyond a skill’s instructions for doing a task. It also needs ways to encode taste in evaluations, run those evaluations, improve them over time and choose useful signals. Tool choices and harness profiles may vary with the model. Provider independence therefore still leaves room for model-specific operating decisions inside the portable recipe.

6:566:58
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:56 · section reference included

Establish value before optimizing its cost

The third idea is “valued work per watt.” Gavrilescu points to Cursor and Cognition as examples of a progression from building a product, to building evaluations for that product, to building models informed by both. He proposes extending this pattern beyond coding into domains such as customer support and legal research. The order matters: first establish what work has value, then determine whether delivering it is a good economic deal.

Source frame: Establish value before optimizing its cost
Source frame: Establish value before optimizing its cost

A base harness and initial evaluations provide a starting point. Running the system in production reveals the requirements needed to reach the desired level of performance. Once the system can produce valuable work, the next research problem is delivering that value without spending more than necessary. “Per watt” serves here as an optimization principle connecting useful output to resource use; the talk does not define an energy measurement or present a numerical efficiency benchmark.

Fine-tuning APIs and infrastructure make parts of this process accessible. The harder know-how is translating judgment into evaluations and validating it through experiments. Gavrilescu describes a route from a maker’s preferences into environments and evaluations, and potentially into model weights. But improving the surrounding system also requires interpreting the artifacts a worker produces and deciding which changes are worth trying.

Experiments then test the maker’s judgment itself. Offline evaluations can establish that an agent satisfies its creator’s criteria. Production use asks whether customers appreciate those criteria. An improvement loop needs both: a way to reproduce the intended judgment and a way to discover whether that judgment creates value for users.

9:189:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:18 · section reference included

A sourcing agent learns that prominence is the wrong signal

The practical example begins with a talent-sourcing agent: web search, LinkedIn, subagents and a system instruction describing its recruiter role. Recruiting makes taste visible because different people leading the process have different ideas of a good candidate. The baseline can search competently while choosing a strategy its owner dislikes.

Source frame: A sourcing agent learns that prominence is the wrong signal
Source frame: A sourcing agent learns that prominence is the wrong signal

Its execution traces reveal the problem: the agent repeatedly reaches out to big-tech employees, while this recruiter wants to find hidden gems. From the agent’s perspective, a prominent candidate looks like an obvious good choice. From the recruiter’s perspective, that choice misses the purpose of sourcing. Grouping similar behaviors and user frustrations into a pattern turns a vague complaint into a specific target for improvement.

The next step is a judge that reads a trajectory—the sequence of actions in a run—and identifies the unwanted pattern. Gavrilescu’s concrete question is whether the sourcing agent contacted Google employees instead of looking for hidden gems on GitHub. That gives subsequent runs a repeatable check. The criterion expresses this recruiter’s strategy, rather than a universal rule that big-tech candidates are unsuitable.

The proposed division of labor puts a human in charge of calibrating that criterion. Does the recruiter agree that the agent should favor hidden gems over big-tech outreach? Agents can build the evaluations and put the judgment into code; the person checks whether they are encoding the intended preference. This is where the observed failure becomes an explicit product decision.

12:3012:32
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:30 · section reference included

Test the recipe with users before promoting it

A calibrated judge makes it possible to create recipe candidates: changes to the system that can be compared with the existing version. In the sourcing example, the intended observable change is a move away from big-tech outreach toward discovering hidden gems. Offline evaluations test whether a candidate follows that preference. Production experiments test whether users value the resulting service. The example describes this improvement process without reporting an experimental result.

Source frame: Test the recipe with users before promoting it
Source frame: Test the recipe with users before promoting it

How does a recurring sourcing mistake become the next recipe version? The diagram follows the example through two distinct judgments: the human confirms the desired recruiting strategy, then users test its value in production. The path makes clear why a candidate that satisfies the maker still needs an experiment before promotion.

Gavrilescu proposes A/B testing, with a multi-arm bandit as another possible experimental approach, to check whether customers appreciate the maker’s taste. Once that preference is validated with users, the candidate becomes the next version of the agent recipe. Repeating the process lets the system reproduce an increasingly well-defined service for other people.

The ending raises the ambition of the second loop: automate more of the maker’s role as a higher-level judge, so improvement agents can apply the same judgment to workers headed for production. Gavrilescu sees the continued accumulation of that judgment, together with workers that self-verify and cooperate, as a defensible basis for a vertical AI company.

That company still has to earn the customer’s choice. The closing economic test is whether the work is valuable and whether its price makes switching from a general-purpose coding agent worthwhile. Introspection’s invitation to vertical SaaS companies and agent labs is to build these research loops around their own products: learn what good means, reproduce it and deliver it at viable cost.

How it fits togetherFrom sourcing traces to a new recipe

Repeated outreach to big-tech employees

Human calibration establishes the intended preference. Production experiments establish whether users value it.

14:2614:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:09 · section reference included

Read the complete timestamped transcript
  1. 0:12

    Hello everyone. How's everyone doing?

  2. 0:17

    Are you guys ready for some more loops?

  3. 0:21

    Yeah. My name is Roland. My co-founder

  4. 0:24

    and I were in this mythical place called

  5. 0:27

    XAI working hard on agent infra and we

  6. 0:30

    realized there's something new that has

  7. 0:33

    to be done in a standalone way. So we

  8. 0:35

    left a few months ago to really figure

  9. 0:38

    out okay what's the next stage of how we

  10. 0:41

    should deploy these always on longunning

  11. 0:43

    horizon tasks. Um and I'm happy to

  12. 0:47

    announce we have a few findings that we

  13. 0:49

    would like to present you. Um, and this

  14. 0:52

    talk it's all about um, how you should

  15. 0:55

    productize these ideas in ways that can

  16. 0:58

    scale with your customers. Um, you've

  17. 1:02

    heard a lot about auto research. Um, we

  18. 1:05

    think there's a blueprint for 2026 and

  19. 1:07

    beyond on how you should think about

  20. 1:08

    auto research and it really comes down

  21. 1:12

    to three ideas.

  22. 1:14

    Let's go through the first one. The loop

  23. 1:17

    is the product.

  24. 1:21

    We're all familiar with this. We've

  25. 1:23

    started with everything goes down to RL

  26. 1:26

    chief for models and how you should

  27. 1:28

    train the model to become better and

  28. 1:30

    better reasoning. We then quickly moved

  29. 1:32

    to harnesses and how the model is a

  30. 1:35

    commodity and it's all about the

  31. 1:36

    harness. And now we're talking about

  32. 1:38

    loops and how you should build these

  33. 1:40

    loops uh and not touch code anymore. But

  34. 1:43

    what does it really mean and why is

  35. 1:45

    everyone saying that?

  36. 1:47

    Do you guys remember clawbot?

  37. 1:50

    That was the original um original name

  38. 1:53

    of what is now now now known as open

  39. 1:55

    claw. And this guy AJ built the first

  40. 2:00

    loop around cloudbot.

  41. 2:03

    What he did was to find a way to talk to

  42. 2:06

    dealers and talk to Reddit users to get

  43. 2:09

    bigger discounts on a car. He followed

  44. 2:12

    these four steps. Um and is really open

  45. 2:15

    call the one that did it. Go on Reddit,

  46. 2:18

    find prices, find inventory, talk to the

  47. 2:22

    dealers,

  48. 2:25

    put dealers head-to-head, and try to

  49. 2:27

    figure out how to make them outbid each

  50. 2:29

    other,

  51. 2:31

    have a verifiable way to know when the

  52. 2:33

    price is right, and then lock in, get

  53. 2:38

    the car, and it worked. Um, probably

  54. 2:41

    this was when all the Mac minis were uh

  55. 2:44

    selling off the shelves, but this was

  56. 2:46

    the first real example of loop is the

  57. 2:49

    product and something that probably

  58. 2:51

    should be a startup at this point. Um,

  59. 2:53

    but we've seen how this became a recipe

  60. 2:57

    for everyone to build loops. But let's

  61. 2:59

    take a step back. Why are we here? Um,

  62. 3:03

    we really think models have been trained

  63. 3:05

    with this loop in mind. And it comes

  64. 3:07

    from this idea of udala loops. It's a

  65. 3:10

    terminology coined back in 1970s by the

  66. 3:15

    US air force and is the idea of these um

  67. 3:19

    jet fighters how to react in fast-paced

  68. 3:22

    environments.

  69. 3:24

    If you think of models calling tools and

  70. 3:27

    taking observations, it's it's what

  71. 3:29

    we've been trained on uh as humans but

  72. 3:32

    also as as agents. Now, now what happens

  73. 3:35

    when you put strong signals and

  74. 3:37

    verifiable work uh at the other ends?

  75. 3:41

    You get to these workers or cloud code

  76. 3:44

    agents. Um and and what matters here is

  77. 3:48

    the quality of the signal determines the

  78. 3:51

    uh success rate of the loop and the uh

  79. 3:55

    quality of the verifi verifier um um is

  80. 3:59

    able to calibrate if that success is

  81. 4:02

    actually correct or not. But there's

  82. 4:05

    another loop here. Um what happens when

  83. 4:07

    you take that and feed it back into the

  84. 4:09

    signal? And this is what looping around

  85. 4:12

    is all about is how do you generate

  86. 4:15

    these artifacts at the end of the first

  87. 4:17

    loop to then run a second loop on and

  88. 4:20

    have a way to continuously improve. And

  89. 4:24

    this goes to my second point. System

  90. 4:26

    dillation is the mode and is really the

  91. 4:30

    ability to understand what went well and

  92. 4:34

    wrong in the first loop and know how to

  93. 4:37

    process that in the second one.

  94. 4:40

    So how do we tune these AI systems? Each

  95. 4:43

    loop generates useful information around

  96. 4:45

    harnesses, profiles, evals,

  97. 4:48

    models, resources, tools, and the

  98. 4:51

    environment. What you really want is

  99. 4:55

    to have a way to keep this portable, to

  100. 4:57

    have a way to version this and to evolve

  101. 4:59

    it over time. If you think about data

  102. 5:03

    recipes in research, this is how RL

  103. 5:07

    started to work really well. you

  104. 5:09

    understood the recipes and how to

  105. 5:11

    continuously change the recipe to combat

  106. 5:14

    some of the behaviors that may happen

  107. 5:15

    around hallucinations around reward

  108. 5:17

    hacking and then you get to a stack

  109. 5:20

    which is your final data recipe. We

  110. 5:22

    don't have that for harnesses. We don't

  111. 5:24

    have that for like AI systems in the

  112. 5:26

    general term. So we thought there's

  113. 5:28

    space for something like that. something

  114. 5:31

    that contains the evils and contains the

  115. 5:33

    tweaks and the human judgment and all

  116. 5:36

    these things that are not predetermined

  117. 5:37

    at the beginning but they're defined as

  118. 5:40

    you learn more about your agent acting

  119. 5:42

    in in in the environment.

  120. 5:47

    We think recipes can be applied to this

  121. 5:49

    and we should use the same name. So an

  122. 5:52

    agent recipe is really something that

  123. 5:54

    enables you to create reproducible

  124. 5:56

    frontier AI systems. It's something that

  125. 5:59

    allows you to have a mode that keeps

  126. 6:02

    getting better over time, which is not

  127. 6:04

    tied to any platform or any provider.

  128. 6:07

    It's something that you control lives in

  129. 6:10

    your company and is agnostic to the

  130. 6:13

    models and providers you use. And loops

  131. 6:16

    should focus on this. Loops should be

  132. 6:18

    the way you distill these systems into

  133. 6:20

    recipes.

  134. 6:21

    Failure patterns should become judges

  135. 6:23

    and evals. Repeated behavior should

  136. 6:26

    become skills and prompts. user

  137. 6:27

    frustration, extensions and memories to

  138. 6:29

    your harness and so on. You we're all

  139. 6:32

    familiar with this, but we didn't have

  140. 6:34

    the the the right like terminology of

  141. 6:36

    how we should think about it and how we

  142. 6:38

    should define it. And we think recipes

  143. 6:40

    is a way to put everything together into

  144. 6:45

    a git repo and treat it as your ongoing

  145. 6:50

    um strategy for for uh building these

  146. 6:53

    self-improving systems. So we are

  147. 6:56

    introspection but you can think of

  148. 6:58

    introspection as the way you generate

  149. 6:59

    these recipes. So they're recipes for

  150. 7:02

    introspecting on your on your system. We

  151. 7:05

    wanted to build something that is

  152. 7:07

    portable and provider agnostic. So we

  153. 7:09

    built our um approach to recipes on the

  154. 7:13

    pi harness and on harbor for evals.

  155. 7:17

    We baked it into uh git repos so uh

  156. 7:21

    everything could be versioned and agents

  157. 7:24

    could have a way to continuously track

  158. 7:26

    how this change and why and is meant to

  159. 7:29

    be owned by you but managed by your

  160. 7:30

    agents. And this is how products should

  161. 7:33

    really be built going forward. It's

  162. 7:34

    something that treats the owner as the

  163. 7:38

    um almost like the the the higher taste

  164. 7:42

    um personality in the room. But agents

  165. 7:44

    should try to calibrate themselves to to

  166. 7:47

    the taste of the of the maker. So we

  167. 7:50

    think recipes should be basically

  168. 7:52

    encoding the taste of the makers into

  169. 7:54

    how you build these agents. And if I

  170. 7:57

    want to use someone else's recipe, I

  171. 7:59

    should be able to also bring that taste.

  172. 8:01

    It's not just the harness, it's not just

  173. 8:03

    the model, is how did you arrive at this

  174. 8:05

    particular recipe and why? And that's

  175. 8:08

    kind of like what uh what is behind uh

  176. 8:12

    reproducible

  177. 8:13

    um uh products and services around

  178. 8:16

    agents. Um we have an early release of

  179. 8:19

    recipes is called pi. Recipes. It's very

  180. 8:23

    similar to what skills uh used to be in

  181. 8:25

    2025 but is going a step forward. And

  182. 8:28

    this is what do I need to have a

  183. 8:30

    frontier agent is everything about how

  184. 8:33

    do I codify paste into evals? How do I

  185. 8:36

    run evals? How do we have the loops to

  186. 8:38

    continuously improve those evals over

  187. 8:40

    time? How do we process signals and know

  188. 8:42

    what are the right signals to to use? Um

  189. 8:45

    what are the right tools to work with

  190. 8:47

    certain models? How do I have different

  191. 8:49

    profiles of the harness to work with

  192. 8:51

    different models? Um and everything in

  193. 8:53

    between. So have a look at what we've

  194. 8:56

    been building here. It's still early uh

  195. 8:59

    but hopefully it's useful enough for you

  196. 9:01

    guys to to get going. And we feel this

  197. 9:03

    is going to grow into something that um

  198. 9:05

    really allows you to to use uh different

  199. 9:08

    um almost like different the to to be

  200. 9:11

    able to use the taste of of different

  201. 9:13

    makers as recipes for your agent.

  202. 9:18

    And finally, the last point is valued

  203. 9:20

    work per watt. And why is this the score

  204. 9:22

    to really optimize for? Think of how um

  205. 9:26

    cursor and cognition went from building

  206. 9:28

    the best product to then building the

  207. 9:31

    best evals for the product and finally

  208. 9:33

    building the best models based on the

  209. 9:36

    previous two artifacts. We think this is

  210. 9:38

    like the recipe for everything going

  211. 9:40

    forward. Um code was the first domain

  212. 9:43

    where this um was successful. Um

  213. 9:46

    everything beyond customer support,

  214. 9:48

    legal research um everything is going to

  215. 9:51

    come down to this idea. How much value

  216. 9:53

    am I getting per watt? Um, how do I

  217. 9:56

    measure the value is the first step and

  218. 9:59

    how do I know I'm getting a good deal on

  219. 10:01

    that value is the second. And maybe this

  220. 10:04

    makes it a bit more clear. We've all

  221. 10:07

    started from a base harness and a base

  222. 10:11

    set of evals and we went to go to the

  223. 10:13

    frontier. Um, and you only go through

  224. 10:15

    that by running these systems in prod.

  225. 10:17

    There's no way you you know what

  226. 10:19

    frontier is before you uh you start. Um

  227. 10:22

    but the the the last step here which is

  228. 10:25

    what is requiring a lot of research um

  229. 10:28

    is okay once you've reached frontier how

  230. 10:30

    do we make this um uh economically

  231. 10:34

    viable which is how do we not spend more

  232. 10:37

    than than uh we need for generating this

  233. 10:39

    amount of value um and we think we have

  234. 10:42

    the building blocks now to make this

  235. 10:45

    accessible and pretty efficient in the

  236. 10:47

    sense of you've seen all these

  237. 10:49

    fine-tuning APIs all the infrastructure

  238. 10:51

    that has been abstracted away for you to

  239. 10:54

    do do this process is just the knowhow

  240. 10:56

    that uh is not there yet and this is

  241. 10:59

    what we we we hope we can like push for

  242. 11:01

    the knowhow for knowing how to codify

  243. 11:04

    taste into evals and how to validate

  244. 11:06

    that in experiments um and you you've

  245. 11:09

    you've heard a lot about evals in

  246. 11:10

    experiments before but you didn't really

  247. 11:12

    think of them of like what are they is

  248. 11:14

    it's not just tests is is really what is

  249. 11:17

    the taste of the creator that agents

  250. 11:20

    should be able to reproduce and

  251. 11:22

    self-improve around. And no one has

  252. 11:25

    thought of how do I make this as

  253. 11:27

    portable enough? How how do I make my

  254. 11:29

    taste as an artist or as a software

  255. 11:33

    developer um something that anyone can

  256. 11:35

    download in their brain and be able to

  257. 11:37

    be a onetoone replica to me? And this is

  258. 11:39

    kind of like what RL is is is about now

  259. 11:42

    is how do we uh turn these um taste

  260. 11:47

    makers into uh environments and evals

  261. 11:50

    around them so then we can move them

  262. 11:52

    into the weights. But um there's more

  263. 11:54

    than that. Um you can think of the

  264. 11:56

    worker as the inner loop and it

  265. 11:58

    generates all these artifacts. But how

  266. 12:00

    you look at the artifacts and know what

  267. 12:03

    to change is the taste. Uh and this is

  268. 12:06

    what creates candidates of what you

  269. 12:09

    should change and how you should adapt

  270. 12:10

    based on that. And experiments is what

  271. 12:13

    how you self-calibrate that okay my

  272. 12:15

    taste is actually validated in

  273. 12:17

    production with users and we make sure

  274. 12:20

    that not only the maker is happy through

  275. 12:22

    the um offline evals but the end users

  276. 12:25

    are happy as well and they agree with

  277. 12:27

    what we consider good.

  278. 12:30

    Let's go through a practical example of

  279. 12:32

    how this works.

  280. 12:36

    Let's take a baseline um agent which

  281. 12:38

    could be a talent sourcing agent. Um and

  282. 12:41

    this is a very classical case of

  283. 12:45

    everyone is doing recruiting differently

  284. 12:47

    and is very much

  285. 12:49

    about not what is good recruiting but

  286. 12:53

    who is leading that recruiting that

  287. 12:55

    considers recruiting as good. So in this

  288. 12:57

    case we're starting with something

  289. 12:58

    pretty simple. um a a bunch of tools,

  290. 13:01

    web search, LinkedIn, uh a bunch of sub

  291. 13:04

    aents that have been pre-popularized by

  292. 13:07

    harnesses like codeex and cloud code and

  293. 13:09

    uh system instruction which is about

  294. 13:11

    your recruiter.

  295. 13:14

    First step is really understand the

  296. 13:16

    signals. So you can think of patterns as

  297. 13:20

    being a way to look at the traces,

  298. 13:22

    extract some common um behaviors or

  299. 13:25

    common user frustrations and turn them

  300. 13:28

    into like a cluster. So let's say this

  301. 13:30

    idea of uh the agent is going uh and

  302. 13:33

    reaching out to a lot of big tech

  303. 13:35

    employees. As a recruiter, you don't

  304. 13:37

    really want that. You want to find

  305. 13:38

    hidden gems. You don't want to try to

  306. 13:40

    hire John Carmarmac. But an agent would

  307. 13:44

    think that's, oh, John Carmarmac is

  308. 13:45

    great. why would I not reach out to him?

  309. 13:47

    Um, so so this is a behavior that you

  310. 13:50

    you'd never think of codifying, but you

  311. 13:53

    discover the agent tends to do that. Um,

  312. 13:57

    patterns is how you discover these

  313. 13:58

    signals and inform you what you should

  314. 14:01

    do next.

  315. 14:03

    um calibration judges and evals is how

  316. 14:07

    we used to think about how do we codify

  317. 14:10

    these these behaviors into um something

  318. 14:13

    that can try to uh apply the same

  319. 14:15

    judgment across traces and across uh

  320. 14:17

    execution. So let's say we we build an

  321. 14:20

    agent that looks at a trajectory and um

  322. 14:24

    identifies exactly that pattern. Hey,

  323. 14:26

    did did this agent reach out to Google

  324. 14:28

    employees instead of trying to uh find

  325. 14:31

    hidden gems on GitHub? Um, and the

  326. 14:34

    calibration bit and the eval generation

  327. 14:36

    bit is not that hard. It it it should be

  328. 14:39

    doable by agents to build. You just need

  329. 14:41

    a human in the loop to say, hey, um,

  330. 14:43

    this is the approach we're taking. Do

  331. 14:45

    you agree with this

  332. 14:48

    judgment? Do you really agree that we

  333. 14:50

    should look more towards hidden gems

  334. 14:52

    rather than reach out to um um big tech

  335. 14:56

    employees? And that's about it. You

  336. 14:58

    don't need the human to actually build

  337. 15:00

    the evals. You need them to calibrate

  338. 15:02

    the evals. And agents should be the ones

  339. 15:05

    that really take the the the taste of

  340. 15:07

    the maker and and put them in into code.

  341. 15:09

    Once you have this, it's pretty easy to

  342. 15:12

    create recipe candidates. And this

  343. 15:13

    should be the the diffs that you really

  344. 15:15

    want to taste. Um, and

  345. 15:19

    you can have a pretty good offline evil

  346. 15:21

    set around this, but the the the test

  347. 15:23

    here is when you go to prod. So, do the

  348. 15:27

    end user agree with your taste of not

  349. 15:30

    hitting up um big tech uh employees,

  350. 15:34

    right? And this is kind of like what you

  351. 15:36

    want is you build a product that really

  352. 15:38

    emphasizes your taste and then you you

  353. 15:40

    make sure that your users appreciate and

  354. 15:43

    value that taste. and AB tests have been

  355. 15:46

    a way to to to make sure that that's the

  356. 15:48

    case. Um so with a multi-arm banded um

  357. 15:51

    scenario for example you you'd be able

  358. 15:53

    to do that pretty well. So once you

  359. 15:55

    validate okay I have great taste and my

  360. 15:57

    users believe uh I have great taste as

  361. 15:59

    well that's when you promote and that's

  362. 16:01

    kind of when you go to to the next

  363. 16:03

    version of an agent recipe. The secret

  364. 16:06

    is you keep doing this over and over

  365. 16:08

    again and you know how to continuously

  366. 16:10

    codify your taste and your um what what

  367. 16:13

    what good is to you into an agent that

  368. 16:16

    can reproduce the same service or

  369. 16:18

    product uh for other people and they

  370. 16:20

    also agree you have great taste and you

  371. 16:23

    have great execution. And this is really

  372. 16:24

    kind of like the the secret of building

  373. 16:26

    good loops is okay can can someone

  374. 16:28

    iterate on my um system in a way as uh

  375. 16:33

    you know um a good example here is like

  376. 16:35

    Miranda from the Delor product right

  377. 16:38

    what would Miranda do uh in certain

  378. 16:40

    cases and you kind of want to codify

  379. 16:42

    that that thinking into like agents that

  380. 16:44

    can do the same stuff at a higher level.

  381. 16:48

    So the takeaways are this. Um the loop

  382. 16:50

    is the product. You try to automate

  383. 16:51

    yourself as the u as a um higher level

  384. 16:56

    judge and you want to make sure your

  385. 16:59

    second loop agents are able to apply the

  386. 17:02

    same judgment to to the agents you're

  387. 17:04

    trying to to to push to prod. Second bit

  388. 17:07

    system dissolation is the mode. So, how

  389. 17:09

    do you continuously inject that taste

  390. 17:11

    into these uh workers and they how how

  391. 17:13

    they continuously self-verify and work

  392. 17:15

    together is uh the biggest thing that

  393. 17:18

    you should focus on and the faster you

  394. 17:20

    do it uh the the the faster you you

  395. 17:23

    build a defensible

  396. 17:25

    um approach to to becoming a vertical AI

  397. 17:27

    company. And finally, valued work per

  398. 17:30

    watt is how you should measure um am I

  399. 17:33

    making progress or not. So first make

  400. 17:36

    sure that uh the the the work you're

  401. 17:38

    generating is valuable. Second make sure

  402. 17:40

    that the economics makes sense and the

  403. 17:44

    um the the difference in price is is

  404. 17:45

    basically what um people would would

  405. 17:48

    switch away from cloud code to to

  406. 17:50

    something you provide.

  407. 17:53

    We've been thinking a lot about these

  408. 17:54

    ideas and we're building some very

  409. 17:56

    interesting products around how to

  410. 17:57

    deploy this in production. We'd love to

  411. 17:59

    hear from you. would love to get um to

  412. 18:02

    to understand more about how how certain

  413. 18:05

    um vertical SAS companies are are

  414. 18:07

    looking to go to prod with um or how

  415. 18:10

    agent labs have been thinking about this

  416. 18:13

    idea of um um creating these like auto

  417. 18:16

    research uh labs around their their own

  418. 18:19

    products. um get in touch. Uh we're

  419. 18:21

    gonna be around the block for for

  420. 18:23

    chatting more about this. And thank you

  421. 18:25

    very much.