I don't prompt agents anymore... — Transcript

Full transcript

1. 0:00Thanks HubSpot for sponsoring this
2. 0:02video. So today I want to talk about
3. 0:04graph engineer. If you're on Twitter,
4. 0:06you probably get pretty confused past
5. 0:07few weeks. It's just new hot term that's
6. 0:10popping up everywhere on internet and
7. 0:11the more you look at it probably the
8. 0:13more confusing it is because people are
9. 0:15mixing multiple different things awkward
10. 0:17graphs into same discussion. So in
11. 0:19today's videos I want to break down what
12. 0:21kind of practical things that you can
13. 0:22take away and start adopting for real.
14. 0:24So it all start from Peter tweet about
15. 0:27are we still talking loops or do we
16. 0:28shift to graph yet and no one really
17. 0:30understand what he meant about graph
18. 0:33because graph can mean just so many
19. 0:35different things and that's where the
20. 0:37conversation became really confusing
21. 0:38because people are basically using graph
22. 0:40as term to discuss three totally
23. 0:43different concepts what is control graph
24. 0:45so control graph you can think about
25. 0:46lang graph dynamic workflow basically
26. 0:49most of those agenda task have some sort
27. 0:51of SOP behind it and control graph like
28. 0:54lang graph or recently dynamic workflow
29. 0:57or even just SOP you put in text are
30. 0:59ways for you to making sure agent get
31. 1:02more reliable result and I think this is
32. 1:04actually the graph that Peter is talking
33. 1:06about because in his recent tweet the
34. 1:08use case he's describing is basically
35. 1:10control flow SOP of shipping a change to
36. 1:12a codebase so this is control flow the
37. 1:15second type of graph people talk about
38. 1:16is knowledge graph and this is a
39. 1:18completely irrelevant and different
40. 1:20concept but just because Andrew Num
41. 1:23happened to released a 1hour course on
42. 1:25building a genetic knowledge graph.
43. 1:26People somehow connecting them together
44. 1:28and knowledge graph is just a totally
45. 1:30different concept. Basic graph is a
46. 1:32greater way to representing the
47. 1:34relationship between different entities.
48. 1:36Which means when people are retrieving
49. 1:38data, it can retrieve information
50. 1:39effectively. That's why gradiki agent
51. 1:43memory that view has been shipping with
52. 1:46graph for years. But it is totally
53. 1:48different topic. The third topic people
54. 1:50were start talking about is actually
55. 1:52very interesting called graph of loops
56. 1:54and this concept is brought out by
57. 1:56colors. If we assume the future company
58. 1:58were run by a bunch of different loops.
59. 2:00How can you making sure all those loops
60. 2:02actually execute well and improve
61. 2:04continuously because even just one loop
62. 2:06is hard and when you have multiple
63. 2:08different loops both error and
64. 2:09improvements can compound very quickly.
65. 2:12And this is a very much new and
66. 2:13unexplored space. And to be honest,
67. 2:15there's no real answer to how do you
68. 2:18solve that because again most of the
69. 2:20team haven't even set up one or two
70. 2:21loops successfully. But my team has been
71. 2:24exploring loops for the past two months.
72. 2:26And we did some experimentation which I
73. 2:28put into one of the past articles that
74. 2:30you can dive up deeper if you're
75. 2:31interested. But those are the three
76. 2:33different topics that people has been
77. 2:35mixing together on the internet when
78. 2:37they talk about graph engineer. But the
79. 2:39real one Peter was referring to, I
80. 2:41believe, is a control graph idea. And
81. 2:43based on where most companies are at, I
82. 2:45think the broad practical usefulness
83. 2:48about this control graph idea. But
84. 2:50before we dive into what the control
85. 2:51graph actually is, everything I'm about
86. 2:53to show you sit on top of modern harness
87. 2:55cloud code or codeex. But there's
88. 2:57probably already a 10 or even 100x
89. 3:00difference between someone who just open
90. 3:02a terminal, sends a prompt, and hopes
91. 3:04for the best versus someone who actually
92. 3:06use every part of what those modern
93. 3:08harness give you. And if you want to be
94. 3:10on the right side of the gap, this free
95. 3:12cloud code playbook I put together with
96. 3:14HubSpot will get you there. It covers
97. 3:16the prompting patterns I use. How do I
98. 3:18structure my cloud MD file and project
99. 3:21context to make the most out of it? the
100. 3:23hooks and guard rails I put together.
101. 3:25How do I keep long running work on track
102. 3:28with task state and when the sub aents
103. 3:30are actually worth it versus when they
104. 3:32are just tripling your bill? A fourth
105. 3:34section on dynamic wolf rolls which is
106. 3:36literally one of the two methods I'm
107. 3:38going to show you later in this video.
108. 3:40And my favorite part is this parallel
109. 3:42session setup that lets you run not just
110. 3:44one or two but 10 to 15 aent sessions
111. 3:47safely at the same time. So, this
112. 3:48playbook I wish I had when I was
113. 3:50figuring everything out the hard way.
114. 3:52I've put the link in the description
115. 3:54below. It's completely free so you can
116. 3:56check out and thanks to HubSpot for
117. 3:58sponsoring this video. Now, let's get
118. 4:00back to the control buff. As I mentioned
119. 4:02before, basically each type of task
120. 4:04normally has certain type of SOP behind
121. 4:06it. And people has this increasing need
122. 4:08to get more reliable outcome from the
123. 4:11agent. And the main driver behind this
124. 4:13needs is because the biggest shift we
125. 4:15are going through now with agent is that
126. 4:17teams are shifting from human prompting
127. 4:20the agent to something else prompting
128. 4:22the agent on behalf of human. For
129. 4:24example, there are two patterns that
130. 4:26some of the most AI native company in
131. 4:28the world are adopting experimenting at
132. 4:30the moment. One is loops and a lot of
133. 4:32people already talk about loops
134. 4:33increment self and basically the idea
135. 4:35here is that instead of you as human to
136. 4:37prompt agent for every single task you
137. 4:40can actually set up a trigger to prompt
138. 4:42agent and the trigger can be time based
139. 4:44like every day go pull the latest GitHub
140. 4:47issue and fix it or it could be go based
141. 4:49like continuously optimize the front end
142. 4:51until the performance increase 200%. Or
143. 4:54it can be event based as well. Every
144. 4:56single email comes in it just wake up
145. 4:58the agent to do the job. And in this
146. 5:00pattern, humor only involved when agent
147. 5:03thinks it's necessary. In fact, one of
148. 5:06the most important thing for each loop
149. 5:08is defining what type of things agent
150. 5:10can do by itself versus one that
151. 5:12actually need to involve human. So this
152. 5:14the first type of pattern that company
153. 5:16are experimenting to get more out of
154. 5:18agent and second type of pattern that's
155. 5:20happening is this ox trader pattern
156. 5:22which means you as a human user no
157. 5:25longer talk to individual agent to ask
158. 5:27to do something instead you talk to
159. 5:30octra agent who has full context about
160. 5:32what's going on and can spin up a teams
161. 5:35of agents to do the actual execution and
162. 5:37also monitoring through the process. In
163. 5:40one video from Boris who is a creator of
164. 5:42cloud code he also mentioned that
165. 5:44nowadays he mostly talked to a cloud who
166. 5:47can talk to other clouds instead of he
167. 5:49prompting the agent for the actual work
168. 5:51directly
169. 5:52>> point where agents are prompting agents
170. 5:54I'm not talking to agents my agents are
171. 5:56talking to other agents and so you have
172. 5:58these kind of big organizations of
173. 5:59agents that that are doing things these
174. 6:01days I think at any given moment I have
175. 6:03you know at minimum a few agents running
176. 6:05often overnight I'll run a few thousand
177. 6:07agents to do some kind of complex work.
178. 6:09Um, this just was not feasible even 3
179. 6:12months ago.
180. 6:12>> As I mentioned, both those two patterns
181. 6:14has been aggressively experimented by
182. 6:16some of the most AI native companies and
183. 6:18my team has been experimenting that for
184. 6:20the past 2 months. What I really
185. 6:22realized is that both two patterns are
186. 6:25keeping you as human one layer above the
187. 6:28actual work. You no longer babysit
188. 6:30through every single step and
189. 6:31checkpoints from the executor agent. And
190. 6:34because of that reliability became very
191. 6:36important for you to have confidence to
192. 6:39adopt such thing. That's why I have been
193. 6:40talking a lot about you have to at least
194. 6:42set up a verifier in your codebase. So
195. 6:45you can have more confidence about what
196. 6:47agent shipping and there are setup
197. 6:49verifier skill that I have included in
198. 6:51the GitHub repo that you guys can go and
199. 6:53take to set up. But verifier just one
200. 6:55small part of the guario for different
201. 6:58type of business processes that you want
202. 7:00agent to automate. Normally there are
203. 7:02different guarios and SOP that can be
204. 7:05injected and this where this control
205. 7:07graph became interesting and this is not
206. 7:09a new concept. This graph based workflow
207. 7:11automation has existed for decades and
208. 7:14normally has three key components. One
209. 7:17is a node which is the action it will
210. 7:19take and the edge which basically what
211. 7:21happen next after each node and third is
212. 7:24state which is what are the data will be
213. 7:26carried over every single node or step.
214. 7:29Even though I said that the concept of
215. 7:31those control graph is not new but
216. 7:34what's new is what does a node actually
217. 7:36means. So in traditional workflow
218. 7:38automation each note is a simple basic
219. 7:40action like a running script do a
220. 7:42conditional check to later like three
221. 7:44years ago platform like allow you to add
222. 7:47a large language model as a node so it
223. 7:50can handle more longtail use cases to
224. 7:52most recently a node can actually be an
225. 7:55agent that's cable to do so many
226. 7:57different things and you're probably
227. 7:59already pretty familiar with graph
228. 8:01because you've been using it for example
229. 8:03the go feature from cloud code or codeex
230. 8:05can be repres present in a graph like
231. 8:07this. At its core is basically this
232. 8:10graph that it receive a prompt from user
233. 8:12which runs through agent loops and when
234. 8:14cloud code or code codex finish it will
235. 8:16get a large learning model to classify
236. 8:19is go satisfied or not. If satisfied
237. 8:21then finish if not send a new message
238. 8:24back to the agent and this graph is
239. 8:26implemented by agent loop plus a stop
240. 8:28hook system. And similarly the famous
241. 8:31project from Andrew Copsy for auto
242. 8:33research is also a graph. If you don't
243. 8:36remember this auto research project is
244. 8:38basically from Andrew Cops. He is
245. 8:40putting agent in a for loop to
246. 8:42continuously improve language model
247. 8:43systems and setup is actually
248. 8:45surprisingly simple. So it also start
249. 8:47with a prompt but this prompt is
250. 8:49documented in a program. file which talk
251. 8:52about the overall processes. It
252. 8:54basically instruct agent to update the
253. 8:56train. PY which is the software script
254. 8:59itself with a experimental idea and then
255. 9:02run the experiment and evaluate the
256. 9:04result. If the result is good then keep
257. 9:06the change. If not discard this and try
258. 9:09a new one. And this fundamentally is a
259. 9:11loop that can be represented with this
260. 9:13specific graph process. Every time it
261. 9:16receive the program.md file to instruct
262. 9:18agent about the processes to try a new
263. 9:20idea and once it finish it will do a
264. 9:23metric check. Is it getting better? If
265. 9:26it is then come in and push a change. If
266. 9:28not discard and revert and also try the
267. 9:31next idea and the state here are
268. 9:33basically two things. One is a train py
269. 9:35which is a software it is changing and
270. 9:37the other is results.tsv that is logging
271. 9:40all the things that you try and
272. 9:42interesting you will see here is that
273. 9:44this is a graph that achieved by simply
274. 9:46prompting the agent. It didn't do any
275. 9:48fancy mechanism. Back the days it didn't
276. 9:50even have rough loop or go feature. It
277. 9:53simply just tell the agent you repeat
278. 9:55this process forever and agent just did
279. 9:58it. And this is kind of interesting part
280. 9:59that I will talk a bit more because
281. 10:01implementation of graph doesn't have to
282. 10:03be code. In fact, it can be just a
283. 10:05documented SOP just as a bunch of
284. 10:07predefined script and boundaries.
285. 10:09Another example is like deep research.
286. 10:11If you ever build agent before, you'll
287. 10:14be pretty familiar with this pattern
288. 10:15that if you are building a research
289. 10:17agent, the common pattern is that it
290. 10:19will receive a prompt. It goes to
291. 10:21octrader who will plan out what type of
292. 10:24topics it should cover. Then fan out a
293. 10:26bunch of sub aents to do research on
294. 10:29interview topics and once it finish it
295. 10:31in to do a review is this enough
296. 10:33information to covering the report. If
297. 10:35no go back to previous step if yes then
298. 10:38just finish and generate report. And
299. 10:41again this is a graph and after going
300. 10:42through those examples hopefully you
301. 10:44realize one thing which is there are
302. 10:47multiple ways you can actually enforce
303. 10:49the graph. One is traditional code as
304. 10:51graph and there are things like dynamic
305. 10:53workflow from cloud code or codec code
306. 10:56mode as well as lang graph that you can
307. 10:58use to build this type of graph workflow
308. 11:00and it is very useful but on the other
309. 11:02hand a lot of time you can just use
310. 11:05large model as graph and many of you
311. 11:07probably already doing it you just
312. 11:08simply include a SOP that can be text or
313. 11:12use JSON or mermaid chat whatever modern
314. 11:15model is already pretty capable of
315. 11:17understanding those SOP workflows and
316. 11:19following through them. So many times
317. 11:21you can actually just describe your
318. 11:23design SOP in text and turn them into a
319. 11:26skill that agent can invoke as context.
320. 11:29That agent will be able to run the graph
321. 11:31pretty well just like Andrew Cupsy's
322. 11:33auto research. And there are a lot of
323. 11:35primitives already provided by the
324. 11:38modern harness like those coding agent
325. 11:40can run bash which means it can run a
326. 11:42predefined script that you define for
327. 11:44those deterministic behavior and steps.
328. 11:47It also has sub agents and an agent
329. 11:49teams communication protocol which means
330. 11:51a main or trader agent can decompose a
331. 11:54SOP into steps and wake up a team of
332. 11:57agents. And if you really want you can
333. 11:59also use hooks to inject real
334. 12:02deterministic checkpoints into the agent
335. 12:04runtime and the goal implementation is
336. 12:06one great example and their practical
337. 12:08usefulness for both method and I'll talk
338. 12:10through them but for the larger model as
339. 12:12graph mechanism it might feels very
340. 12:15basic and simple you basically just give
341. 12:17it a prompt but there are some design
342. 12:19patterns that I think is quite useful
343. 12:21and important to making sure the SOP are
344. 12:23actually well followed by the agent. One
345. 12:25is that you should know when to separate
346. 12:27out agent nodes because technically you
347. 12:30can put all the SOP into just one agent
348. 12:32and as the model gets better more likely
349. 12:35you can just finish the host thing but
350. 12:36it's also important to understand the
351. 12:38boundary for the model so you can making
352. 12:40sure you get the best performance for
353. 12:41large model no matter how powerful it is
354. 12:44not that great at selfverify its own
355. 12:46work that's why a verifier is normally a
356. 12:49separate agent node and also based on
357. 12:51the research from open AI and entropic
358. 12:54normally For complex task, a dedicated
359. 12:57planner agent who is just focused on
360. 12:59researching, reasoning, thinking things
361. 13:01through are necessary. And on the other
362. 13:03hand, you should probably be thinking
363. 13:05about what are things that can be
364. 13:06parallelled and what are things that can
365. 13:08be potentially handled by stronger or
366. 13:10weaker models. Those are all things that
367. 13:12you can think about when you're planning
368. 13:14what type of nodes and steps should
369. 13:16actually exist. And second point that
370. 13:18you should really try to use code when
371. 13:21things make sense. This is one great way
372. 13:23to increase reliability and also making
373. 13:26sure your agent works faster makes less
374. 13:28unnecessary mistakes and common scenario
375. 13:30that I think code actually should apply
376. 13:32are workflows where you might have
377. 13:34complex data fetching. So instead of
378. 13:36getting agent to call multiple different
379. 13:38API putting together some data, you can
380. 13:40have scripts for those complex but
381. 13:42common data analysis and data fetching
382. 13:44and on the other hand you probably
383. 13:46should have script to just get dev
384. 13:47server up running so your agent don't
385. 13:49need to fight about it as well as the
386. 13:51ability to evaluate the result or do
387. 13:54some end to end tests. And third one is
388. 13:56that when you have agent node you
389. 13:57typically want to define what's the
390. 13:58input and output from that agent because
391. 14:01you want to making sure the boundary and
392. 14:02output expectation from each node is
393. 14:05very clear. And fourth is a state. So no
394. 14:07matter whether you use large mode as
395. 14:08graph or code as graph, you need a way
396. 14:10for different agents to get an
397. 14:12understanding of the current state of
398. 14:13the work very quickly. And there are
399. 14:15many different ways you can do it. The
400. 14:16easiest way is that you can just have
401. 14:18one markdown file that's documenting
402. 14:20what's the latest status as well as a
403. 14:22pen only lock from each agent or each
404. 14:25run. And there probably more thing to it
405. 14:27and you guys can let me know if there
406. 14:28are some stuff I missed. But in general
407. 14:30the process is that for any task that
408. 14:32you want agent to repeat or reliably
409. 14:35doing you basically just map out the SOP
410. 14:37think through all steps and after that
411. 14:39you start grouping things into different
412. 14:41nodes and edges with a design pattern
413. 14:43that we mentioned here and in the end
414. 14:44you will need to define an artifact dock
415. 14:47that can document state and this process
416. 14:49I think apply for either code as graph
417. 14:51or larger model graph and the only
418. 14:53difference is output the larger model as
419. 14:55graph approach the output can be simple
420. 14:57skill that can be wake up by Whoop run
421. 14:59or ad hoc versus code as graph will be
422. 15:02operating JavaScript that can be run and
423. 15:04let's just talk through some real
424. 15:05examples. So I want to show you how my
425. 15:07team are adopting these two patterns
426. 15:10that I mentioned before and then show
427. 15:11you two example each how we distain type
428. 15:15of tasks. So many of you probably know
429. 15:17my team built this uh super design which
430. 15:17team built this uh super design which is
431. 15:19is a vibe design platform that allow
432. 15:19a vibe design platform that allow people
433. 15:21people to design user interface and we
434. 15:21to design user interface and we have
435. 15:23have been running multiple different
436. 15:23been running multiple different loops
437. 15:24loops across this business and the way
438. 15:25across this business and the way we were
439. 15:27we were doing that is we actually have a
440. 15:27doing that is we actually have a
441. 15:29separate uh ripple we call super design
442. 15:29separate uh ripple we call super design
443. 15:31AGI this is used as a state tracking and
444. 15:31AGI this is used as a state tracking and
445. 15:34knowledge base for everything that's
446. 15:34knowledge base for everything that's
447. 15:36happening across the business and
448. 15:36happening across the business and
449. 15:38nowadays when I try to make a change to
450. 15:38nowadays when I try to make a change to
451. 15:41my codebase instead of going to the
452. 15:41my codebase instead of going to the
453. 15:43actual ripple which is a super design
454. 15:43actual ripple which is a super design
455. 15:45platform So many of you probably know my
456. 15:45platform form most time I just start
457. 15:46agent in this ripple which it has a
458. 15:49context about everything that ever
459. 15:51happened for super dai like every we did
460. 15:54every support ticket ever handled or the
461. 15:56engineer ticket is shipped and below it
462. 15:58looks a bit messy but that's kind of
463. 15:59like our artifact stores and at the top
464. 16:02it also have bunch of skills and some of
465. 16:04them are like a small graph that we
466. 16:06build through skills one example is this
467. 16:09um daily bad design triage skew so the
468. 16:12goal here is that every day we have a
469. 16:14bunch of users was creating design on
470. 16:16the platform and we want to capture some
471. 16:18bad design that agent did or complaints
472. 16:20people have so that we can continuously
473. 16:23collecting a data set that we can use to
474. 16:25improve the agent. So we build a loop to
475. 16:27review all the design user generated
476. 16:29every day pick up the ones that's badl
477. 16:31looking or didn't meet users expectation
478. 16:34and add them to list for us to review
479. 16:36and add to our evol data set. So the SOP
480. 16:38or graph we build for this processes is
481. 16:41something like this. So once a day we
482. 16:44will want agent to filter what are the
483. 16:46designs that might be bad and we have a
484. 16:48like a list of heristic indicator that
485. 16:51we have. So the agent can pull raw data
486. 16:53from our database about list of
487. 16:55potential candidates for bad design. If
488. 16:57there's no result then just finish the
489. 16:59term but if there are potential bad
490. 17:01designs at that day it will go through
491. 17:03this graph process. Firstly, we create a
492. 17:05script that can help us surface those
493. 17:08issues that can be the heroristically
494. 17:10evaluated things like if there are any
495. 17:12errors or the design just looks totally
496. 17:14broken, not responsive and even some
497. 17:16script that is not possible. And for
498. 17:18those issues, it just directly runs
499. 17:20through the script and add to our list.
500. 17:22But for the other designs, we were
501. 17:24spinning up a bunch of sub agents and
502. 17:26each agent will handle a batch of work
503. 17:28to take a screenshot of design, compare
504. 17:30with the real user request and eval
505. 17:33output result. And with all the
506. 17:35information gathered, the agent will
507. 17:37rank all the bad design and publish to a
508. 17:40today's list which is a simple file in
509. 17:42this ripple. So we have visibility. We
510. 17:44can add to that evaluation set or spin
511. 17:46up another group of agent for each issue
512. 17:48to improve the agent itself. So for this
513. 17:50graph the way we enforce it is a skew
514. 17:53that we created. We basically build a
515. 17:54skew for this alongside a few different
516. 17:57scripts that we wrote the agent can use
517. 17:59to do some deterministic steps. For
518. 18:01example in the skills we document the
519. 18:03procedure it talk about what script they
520. 18:04should run to get the candidate or bad
521. 18:06design and then instruct agent to run
522. 18:08some determinist script. They can do
523. 18:10some programmatic check to surface
524. 18:12issues early on. then fan out a bunch of
525. 18:14agents for vision logic model judge and
526. 18:17each agent has agreed schema for the
527. 18:19output. Then main agent will rank the
528. 18:22worst of day and add to the report and
529. 18:24we have a output contract for the whole
530. 18:26loop as well as some guar that we put in
531. 18:29place. So with this one every day we can
532. 18:31just start a loop and point agent to the
533. 18:33skus and scripts so it knows what step
534. 18:35it should do and spin up sub agents when
535. 18:37it's needed. And this is one example
536. 18:39implementation of large learning model
537. 18:41as a graph. As you can see, it's
538. 18:42actually pretty complicated workflow,
539. 18:44but we can simply achieve a really good
540. 18:46output by injecting skill plazo scripts.
541. 18:49And on the other hand, we also do use
542. 18:51dynamic workflow as code as graph
543. 18:53method. And if you're not familiar with
544. 18:55dynamic workflow, so agent can basically
545. 18:57write a JavaScript because it was given
546. 18:59a few new primitives and APIs which
547. 19:02specifically are agent. So you can just
548. 19:04call a function to spawn up agent as
549. 19:06well as pipeline. So pipeline means
550. 19:08there are say dependencies between each
551. 19:09stage and also parallel which means you
552. 19:12fan out a bunch of agents. So with this
553. 19:14one agent can basically just write any
554. 19:16javascript by calling those primitive
555. 19:19functions to build any type of graph
556. 19:20that it wants. One example my team has
557. 19:23is this ship change workflow and we
558. 19:25trigger this workflow after every time
559. 19:26we have pretty well definfined scope and
560. 19:29plan for a certain things. In this
561. 19:30workflow it break down into three
562. 19:32different phases. One is the setup and
563. 19:34implement another is verify the work and
564. 19:36if the verification passed it will try
565. 19:38to get agent to simplify and PR the work
566. 19:41inside the javascript file it will
567. 19:43define different schema. So the way
568. 19:45dynamic workflow works or the agent
569. 19:47function works is that you can just spin
570. 19:49up in session with this prompt and
571. 19:51define the model and schema so that when
572. 19:53this agent session finish it will
573. 19:55actually output the result in this
574. 19:56specific schema which is kind of like a
575. 19:59state output from each agent node. The
576. 20:01benefit of this is then you can easily
577. 20:03construct the prompt for next agent node
578. 20:05with important context and details from
579. 20:07the last agent sessions. For example,
580. 20:09you can build a um basic workflow for
581. 20:12let's say writing articles that has two
582. 20:14stage. One is outline article and second
583. 20:16is write article. And for each stage the
584. 20:19code were just to wake up the agent with
585. 20:21this basic prompt and also the schema
586. 20:24they define above which is outline
587. 20:27schema including title bullet points for
588. 20:30different sections. Then phase two could
589. 20:31be write article based on outline. So
590. 20:34you can just write a prompt by
591. 20:36referencing the output from the previous
592. 20:38agent session to construct a prompt and
593. 20:40in the end this function will return
594. 20:42rows information and there are more
595. 20:44things to this dynamic workflow feature
596. 20:46and this actually open source project
597. 20:47called de workflow that they implement
598. 20:49the dynamic workflow in open source way
599. 20:51which means each agent node can be cloud
600. 20:53codeex or any other agent library and I
601. 20:55have few code example they are including
602. 20:58the AI builder cloud community so you
603. 21:00can go and check out if you're
604. 21:01interested so in our super design ai
605. 21:03ripple We were using dynamic workflow to
606. 21:05achieve pretty reliable code shipping
607. 21:08graph. But one thing I want to just
608. 21:10emphasize again is just breaking down
609. 21:13task like this won't work. You actually
610. 21:15want to give agent tools so it can test
611. 21:17things properly. And this is probably
612. 21:19one of the biggest missing step I saw
613. 21:20most people are not doing and wondering
614. 21:22why their loops didn't work. So I would
615. 21:24highly recommend you go set up if you
616. 21:27don't know how to set up. I actually
617. 21:28have this verifier setup skew in our
618. 21:31public ripple from AI buildup. So you
619. 21:33can just go and grab it. It's just going
620. 21:35to unlock a lot of other things that you
621. 21:36can possibly do. So by showing those two
622. 21:39example hopefully now you can see how
623. 21:41you can actually do graph engineer and
624. 21:43there are two method you can use. Each
625. 21:45one of them has different output and
626. 21:47both are valid. I personally normally
627. 21:49will go with large motor as graph method
628. 21:51in most of the cases because with
629. 21:52dynamic workflow or codeex code at
630. 21:55moment every time you just start a new
631. 21:56agent session so it can't really resume
632. 21:59the previous conversation and that made
633. 22:01it see things a bit more annoying but if
634. 22:03your task is extremely big code as graph
635. 22:05is actually really good method and in
636. 22:07next video I actually want to talk about
637. 22:09this concept of graph of loops what
638. 22:11happen when you have so many loops
639. 22:13running your company how can you making
640. 22:15sure they compound so if you're
641. 22:16interested Please comment, subscribe.
642. 22:18Uh, and I hope you enjoy this video.