YouTube2Text

Arrêtez les prompts. Passez aux Loops (Le futur des agents IA) — Transcript

by Elliott Pierret · 7,430 words · 1,081 segments · language en · Watch on YouTube

Full transcript

  1. 0:00You need to stop creating agents, and
  2. 0:02it’s not just me saying it; it’s
  3. 0:04Boris Chouhy, the creator of Claude
  4. 0:06Code, and Peter Steinberger, the
  5. 0:08creator of OpenClow, saying it. And for
  6. 0:10them, they summarized it in one
  7. 0:11sentence. They no longer write prompts
  8. 0:13in Claude. Instead, they write loops
  9. 0:16that prompt for them in Claude and
  10. 0:18decide which actions to take. Basically
  11. 0:21, their new job is to write loops. But
  12. 0:23the problem is that the word "loop"
  13. 0:24doesn't mean anything to most people.
  14. 0:26What is a loop used for? What does it
  15. 0:28look like? When should we create one,
  16. 0:30and most importantly, how is it
  17. 0:31actually built? Well, in this video,
  18. 0:33I’ll clear all of that up. I will
  19. 0:35define what a loop is and what you need
  20. 0:36to put inside it. I will also show you
  21. 0:39many use cases for loops, but more
  22. 0:41importantly, I’ll provide concrete
  23. 0:43demonstrations for Nike and Trade
  24. 0:44Republic, where I’ll create loops
  25. 0:46that will help you understand the true
  26. 0:48value of this new technique that will
  27. 0:50totally change the way you work with AI
  28. 0:52. Let’s get started creating your
  29. 0:55first loops. So, first, to really
  30. 0:56understand this concept of a loop, we
  31. 0:58need to look back together at the AI
  32. 1:00generations we’ve used so far. If you
  33. 1:03look closely, as we’ve done on the
  34. 1:05channel, we started by trying to
  35. 1:07generate better answers using the AIs
  36. 1:09we use. At the beginning, the model
  37. 1:12produced text based on a prompt, and in
  38. 1:14the end, we had an answer, and all the
  39. 1:16value resided in the quality of that
  40. 1:18response. And you, well, you were
  41. 1:21responsible for writing the right
  42. 1:23prompts to increase the quality of that
  43. 1:25response, and then came the age of
  44. 1:27agents. And there, the goal was no
  45. 1:30longer just to write good prompts, but
  46. 1:32to create systems capable of acting
  47. 1:34perfectly based on a request. Then our
  48. 1:37agent would call tools, it had memory,
  49. 1:39and in the end, it was capable of
  50. 1:41performing actions for us. We set goals
  51. 1:43, provided a work environment, and in
  52. 1:45the end, we had a result. But now, with
  53. 1:48this notion of a loop, we’re arriving
  54. 1:50at a whole new generation of agents.
  55. 1:52The idea is to build systems capable of
  56. 1:54learning on their own. Ultimately, we
  57. 1:57give it a framework within which it can
  58. 1:59execute. Then, it performs tasks. It
  59. 2:02then evaluates the result itself or
  60. 2:04with our help. It will then correct
  61. 2:07itself, and in the end, it will
  62. 2:08understand the adopted behaviors so
  63. 2:09well that it will be able to do it
  64. 2:11alone without our intervention. Over
  65. 2:13the course of executions, it will
  66. 2:15improve, and the result will always be
  67. 2:17of the same or better quality. If you
  68. 2:19look, you can see it even better here.
  69. 2:21At first, we typed prompts, called the
  70. 2:23model, and got the answer. Now, we have
  71. 2:25a system that will be able to produce
  72. 2:28actions. What kind of actions? Well,
  73. 2:30here, the actions we will have defined
  74. 2:32ourselves under our supervision where
  75. 2:34we gave the trajectory, it will produce
  76. 2:36actions. Then, it will evaluate the gap
  77. 2:39between what it provided and what we
  78. 2:41initially wanted. And, in the end, it
  79. 2:44will correct itself to ensure that the
  80. 2:46next time it produces an action, it
  81. 2:49will be of good quality, always with
  82. 2:51supervision from us, but also its own
  83. 2:53supervision and control within the
  84. 2:55system. So what you need to remember
  85. 2:59here is that when we create loops in
  86. 3:01this new generation of AI, the goal
  87. 3:02will no longer be to provide a prompt
  88. 3:04and get a result, but to explain
  89. 3:06everything that happens inside a system
  90. 3:08. And this system, in the end, must be
  91. 3:11able to run without you, and you just
  92. 3:13have to watch it, supervise it, and
  93. 3:14then come in to rectify it if there are
  94. 3:16occasional discrepancies. But otherwise
  95. 3:19, the system is self-learning,
  96. 3:21self-correcting, and always provides
  97. 3:23you with the results you need at
  98. 3:24certain times, under certain conditions
  99. 3:26. And so, once we've said that, I
  100. 3:28suggest we now go back to the loop
  101. 3:30architecture. What do we need to put
  102. 3:32into these loops for them to work? If
  103. 3:35you look here, it comes from an
  104. 3:37analysis done by the Anthropic teams,
  105. 3:39who share the best possible
  106. 3:40architecture for what we call a
  107. 3:42recursive loop. One that is capable of
  108. 3:45never stopping and having a continuous
  109. 3:47cycle. We will have a planner that
  110. 3:49breaks the objective down into a
  111. 3:51realistic plan. It ultimately provides
  112. 3:54the plan and the steps to accomplish,
  113. 3:55and above all, it sets the course with
  114. 3:57the final goal. Next, we have the
  115. 4:00builder. It applies the plan and calls
  116. 4:03the right tools at the right time. You
  117. 4:05need to send an email, it calls Gmail,
  118. 4:07view data in the CRM, or retrieve a
  119. 4:09customer record from your ERP. It is
  120. 4:12the one that implements the plan and
  121. 4:13calls the right tools. Next, well,
  122. 4:15within this plan, there may be some
  123. 4:16creation, some elements to be done.
  124. 4:18Ultimately, the bulk of the work is
  125. 4:20done by the builder. Then, once it's
  126. 4:23finished, we'll put a verifier in the
  127. 4:25loop to ensure that the output meets
  128. 4:27our initial expectations. Once we have
  129. 4:30that, there's a super important role in
  130. 4:32agents since their inception, which is
  131. 4:34memory. How do we ensure that within
  132. 4:37the system, we have a memory that can
  133. 4:39understand successes, but also failures
  134. 4:42, and thus self-improve the elements
  135. 4:44within the loop. Next, we'll see the
  136. 4:46manager who will ultimately review the
  137. 4:48plan and see if everything has been
  138. 4:50done or not. And at the end, we have
  139. 4:52the controller. The controller is
  140. 4:53extremely important because it reviews
  141. 4:55the entire loop method. It will go into
  142. 4:57the memory to see, ultimately, were
  143. 4:59there failures, were there successes?
  144. 5:01If so, were there missing elements that
  145. 5:03led to a failure? Were there elements
  146. 5:06we added that were extremely relevant
  147. 5:08and led to a success? All of this
  148. 5:11consists of elements inside the loop.
  149. 5:13And when the controller is finished,
  150. 5:14the goal is not for the loop to stop.
  151. 5:17The controller takes all this
  152. 5:19information, gives it to the planner,
  153. 5:21who, when breaking down the objective
  154. 5:23into a future plan with actions to
  155. 5:25perform, will have taken all of this
  156. 5:27into account, and normally, the chain
  157. 5:29is self-improved, so that in the end,
  158. 5:31we can ensure we always have the best
  159. 5:33result possible. If you look here, I've
  160. 5:36provided a more detailed view of these
  161. 5:38different steps where the planner
  162. 5:40breaks down the objective into a
  163. 5:41realistic plan. The builder applies the
  164. 5:44plan, calls the tools, and produces the
  165. 5:46results. The verifier checks what has
  166. 5:48been produced. Memory keeps track of
  167. 5:50everything that worked or didn't work.
  168. 5:53The manager classifies what remains to
  169. 5:55be done and redirects the subsequent
  170. 5:57operations. And the controller reviews
  171. 5:59the entire method then passes the baton
  172. 6:02to the planner to rerun the whole loop.
  173. 6:04And so, all this is the theoretical
  174. 6:06part of building loops. But you're
  175. 6:08probably wondering how all this
  176. 6:10materializes. Well, right here I've
  177. 6:12laid out the building blocks of a
  178. 6:14functional loop. If you look at this
  179. 6:16loop, it picks up a project from
  180. 6:18Encloud here. We could have a loop as
  181. 6:21shown here, and inside this loop, there
  182. 6:23are key elements that need to be found
  183. 6:25every time. And regardless of whether
  184. 6:28you work in the Cloud or at OpenAI, we
  185. 6:30find the same elements. The first
  186. 6:32concept is the notion of a habit.
  187. 6:34Without it, the loop won't start, and
  188. 6:36you would need human intervention. What
  189. 6:39we want is for it to run automatically.
  190. 6:41What can trigger a loop can be a
  191. 6:43schedule. Every Monday at 9 a.m., it
  192. 6:46runs, or once a month, or once a year.
  193. 6:49For example, every Monday at 9 a.m. is
  194. 6:51what would trigger a loop to perform
  195. 6:53reporting actions. Another example is a
  196. 6:56goal. A goal is an objective we set
  197. 7:00within the loop that continues to run
  198. 7:02until it succeeds. In that case, we are
  199. 7:05looking at long-term projects. Then,
  200. 7:07you have the hooks. Hooks are what
  201. 7:10trigger commands at specific moments.
  202. 7:12For instance, a follow-up cycle could
  203. 7:15trigger a loop. If you want to follow
  204. 7:17up with your clients every 7 days or
  205. 7:19when you haven't received a reply after
  206. 7:21a certain time, the loop triggers to
  207. 7:23contact the client. Imagine this for
  208. 7:25customer support cases as well.
  209. 7:27Whenever you receive a new support
  210. 7:29ticket, the loop triggers from a
  211. 7:31solution and performs all the tasks for
  212. 7:34you. Another loop system could be that
  213. 7:37every time you create a quote and save
  214. 7:39it to your drive, it triggers reporting
  215. 7:41actions, for example. These are all
  216. 7:44technical elements, but they send a
  217. 7:46signal to trigger the loop and perform
  218. 7:48various actions. Here, the loop has
  219. 7:50been triggered. Next, we find things
  220. 7:53like skills inside. This is a concept
  221. 7:55you should normally be familiar with in
  222. 7:57AI. Skills are ultimately how you set
  223. 8:00rules that, once written, can be
  224. 8:02replayed by the AI. I always want you
  225. 8:05to write follow-up emails in this
  226. 8:06format. If you need to generate a quote
  227. 8:09automatically, here is how our quotes
  228. 8:11are drafted. If you are doing reporting
  229. 8:13, here is the slide template to adopt.
  230. 8:16Ultimately, these are all the elements
  231. 8:18within the loop that define how work is
  232. 8:20performed inside the loop. And for that
  233. 8:23, we will rely on the skills we have
  234. 8:25trained beforehand. Next, inside the
  235. 8:27loop, there will be one or more agents.
  236. 8:30Generally, as you'll see, we have a
  237. 8:32central agent at the top who then
  238. 8:34delegates tasks to sub-agents. The idea
  239. 8:37is that we divide the tasks and
  240. 8:38therefore the work among the agents.
  241. 8:41There will be the one who produces, the
  242. 8:42one who validates, the one who plans.
  243. 8:45This allows us to avoid hallucinations
  244. 8:47and ensure the work is well distributed
  245. 8:49, so each agent can call various skills
  246. 8:51and connectors. Speaking of connectors,
  247. 8:55you will obviously put tools and
  248. 8:57connections inside your loop. Tools
  249. 9:00that will allow you to broaden the
  250. 9:03context and the scope of your loop. You
  251. 9:05want to send an email, well, you'll add
  252. 9:07the Gmail connector. You want to be
  253. 9:09able to see your customer data in the
  254. 9:10CRM, you'll add Salesforce. You will
  255. 9:12add all the connectors or MCP servers
  256. 9:15that will come inside the loop at the
  257. 9:17level of the different agents to call
  258. 9:19your context and execute a specific
  259. 9:21task using skills. And finally, as I
  260. 9:24mentioned previously, there is this
  261. 9:26notion of memory which is extremely
  262. 9:28important. Why? Because as the loop
  263. 9:31runs, well, the memory will expand. The
  264. 9:35memory will allow you to specify how
  265. 9:37you want certain things, but above all,
  266. 9:39the memory will recover everything that
  267. 9:41happened in the loop each time, to
  268. 9:43ensure that, for example, if a
  269. 9:45sub-agent produced erroneous results
  270. 9:47two or three times, it can self-modify
  271. 9:49following an inspection by another
  272. 9:51agent to improve performance, without
  273. 9:53you or me having to intervene in the
  274. 9:55loop to manually modify one of the
  275. 9:57agents. The agents will talk to each
  276. 10:00other. One will see that another
  277. 10:02performed poorly, and as time goes on,
  278. 10:05with the help of memory, we will
  279. 10:07understand why and who failed in order
  280. 10:10to self-improve. So now you should
  281. 10:12normally understand what a loop is and
  282. 10:14the elements we need to put inside it.
  283. 10:15But then there is a fundamental
  284. 10:17question. Should I leave the loop open
  285. 10:20or should I close it? And for that,
  286. 10:22right here, I have given you two
  287. 10:23examples. Open loops are loops that I
  288. 10:25recommend you use for scenarios, for
  289. 10:28example, aimed at the general public.
  290. 10:30You'll put them on your websites with
  291. 10:32an agent that responds to emails or an
  292. 10:34agent that processes incoming requests
  293. 10:35on a landing page. The idea is that the
  294. 10:38agent will have a very, very wide
  295. 10:39search space. The process has so many
  296. 10:42possibilities that we need to have the
  297. 10:44widest possible loop to handle requests
  298. 10:46. The agent will wander around inside
  299. 10:48and try to find the best possible
  300. 10:50result. However, expect a high
  301. 10:52consumption of tokens at that point.
  302. 10:55The idea here is more for big scenarios
  303. 10:56where the ROI is so strong that the
  304. 10:58budget is almost unlimited. So, we can
  305. 11:00afford to have this learning phase at
  306. 11:02the beginning. Be careful, just because
  307. 11:04the loop is open and very wide doesn't
  308. 11:06mean it will always stay that way.
  309. 11:08Ultimately, the idea of the loop, once
  310. 11:10again, is that it should be
  311. 11:12self-learning, so after a while, we can
  312. 11:14imagine that inside the loop, the field
  313. 11:16of possibilities narrows to provide
  314. 11:18much higher-quality answers. The
  315. 11:20concept is fewer barriers but more
  316. 11:22creativity. The closed loop is really
  317. 11:24what I recommend you start with. The
  318. 11:27idea is that it will be a closed
  319. 11:28process where you provide the different
  320. 11:30steps to be completed. That is exactly
  321. 11:33what we will create together in the
  322. 11:34demo shortly. The idea inside is that
  323. 11:37we will have a clear evaluation of each
  324. 11:39step. We will have a more restricted
  325. 11:41budget because it will consume fewer
  326. 11:43tokens. But most importantly, your loop
  327. 11:45is very structured. You know the final
  328. 11:47result you are looking for, so
  329. 11:49naturally, you will create the loop
  330. 11:50accordingly, starting from what you
  331. 11:52want to achieve and including all the
  332. 11:54necessary elements in the loop every
  333. 11:55time. So, at the beginning, stick more
  334. 11:57to closed loops, but know that you can
  335. 11:59also move on to open loops depending on
  336. 12:01your scenarios. And then you might ask
  337. 12:03me, "OK great, but at what point should
  338. 12:05I actually create a loop?""Why can't I
  339. 12:07just keep using an agent or just chat
  340. 12:09with the AI?" Well, for that, I’ve
  341. 12:11created a little checklist I'm going to
  342. 12:13share with you, which will help you
  343. 12:15understand if your scenario works with
  344. 12:17a loop or not at all. For that, right
  345. 12:19here, you have four conditions that I
  346. 12:21recommend you apply to know if a loop
  347. 12:24is relevant or not. The first condition
  348. 12:26is, does your task repeat itself
  349. 12:27multiple times? Because you need to do
  350. 12:30it every day or every week. Based on
  351. 12:32that, it is necessarily considered
  352. 12:34recurring. So from there, it is a good
  353. 12:36condition to start creating loops. Next
  354. 12:38, the second condition is, is there a
  355. 12:40possibility of failure? Here, the idea
  356. 12:43is: are you able to clearly identify
  357. 12:45when it worked and when it didn't work?
  358. 12:48To do that, if you're not able to tell
  359. 12:50yourself "this is a failure, this is a
  360. 12:52success" every time in a very binary
  361. 12:54way, then there’s really no point in
  362. 12:56using a loop. Why? Because in the end,
  363. 12:59you will always have to go and verify
  364. 13:00the final result. So, there’s an
  365. 13:02intervention on your part. You aren't
  366. 13:04fully in supervision and delegation
  367. 13:06mode. No, you have to come and see all
  368. 13:08the results every time and check them.
  369. 13:11That’s not the goal of a loop at all.
  370. 13:12So, if that’s your situation,
  371. 13:13definitely go with a different scenario
  372. 13:15. However, if you are able to identify
  373. 13:17"this works, that doesn't," and the
  374. 13:19loop can do it for you without
  375. 13:21supervision, then you have a scenario
  376. 13:23for a loop. Next, can the loop do all
  377. 13:25the work, or do you have to come in at
  378. 13:27some point to press a button? Uh, do
  379. 13:29you have to enter text manually at some
  380. 13:31point because you're the one who has
  381. 13:32the information in your head, for
  382. 13:34example? Well here, it's super
  383. 13:35important; the goal is for the entire
  384. 13:37loop to function end-to-end without you
  385. 13:39, and in that case, you have another
  386. 13:41condition met to create your loop. And
  387. 13:44finally, the last condition is: are you
  388. 13:46able to say, "Okay, this is finished,
  389. 13:48there is no human interpretation or
  390. 13:50judgment in what was done"? And this is
  391. 13:53certainly where you will have the most
  392. 13:54difficulty identifying a loop. But,
  393. 13:56you’ll see that we still arrive at
  394. 13:57things where, after a while, we’ve
  395. 13:59run so many tests. It’s on this phase
  396. 14:01, by the way, that the testing part is
  397. 14:03super important. You’ll test the
  398. 14:05results 10, 15, 20 times. After a while
  399. 14:08, if at the end of these 15 or 20
  400. 14:09results you aren't able to say, "It's
  401. 14:11good, all 20 are correct."...then from
  402. 14:13that point on, it means there is human
  403. 14:14judgment involved. There is too much
  404. 14:16interpretation for it to be done by a
  405. 14:18loop, and a human will always have to
  406. 14:19verify it at the end. And in that case,
  407. 14:21it’s not necessarily the goal of a
  408. 14:22loop. If you tell yourself, "Okay, it
  409. 14:24doesn't need me anymore." You have the
  410. 14:25conditions and you can create your loop
  411. 14:27. So now, you have everything you need
  412. 14:28to know when to create a loop. But to
  413. 14:30create one, you need the right tools.
  414. 14:32For this today, I have chosen two tools
  415. 14:34: Codex and Cloud Code. In my opinion,
  416. 14:36they are the two best tools today for
  417. 14:37creating loops. Later, in the demo
  418. 14:39section, we will create loops inside
  419. 14:41Cloud with Cowork as well as with Cloud
  420. 14:43Code. But I still wanted to go over the
  421. 14:46elements of a loop with you. Ultimately
  422. 14:48, what do they look like inside these
  423. 14:49solutions? What building blocks will
  424. 14:51you need to use in these solutions to
  425. 14:53create your loop? And for that, I’ve
  426. 14:55made a small table here with the
  427. 14:56different elements for automation. So,
  428. 14:59what will allow you to trigger, for
  429. 15:01example, the Heartbeat aspect I
  430. 15:02mentioned earlier? Well, for that, in
  431. 15:04Codex you have a section called
  432. 15:06automation which is available in the
  433. 15:08app. You just have to go there. You
  434. 15:10will be able to create small loops
  435. 15:11inside and set goals. Otherwise, on the
  436. 15:14Cloud Code side, as we will see, you
  437. 15:17can use/loop or/g to launch the various
  438. 15:19elements. There is also the whole hook
  439. 15:22part that will allow you to trigger
  440. 15:24from a behavior in an application, for
  441. 15:26example. Next, you will have the work
  442. 15:28areas. This is what will allow you to
  443. 15:30distribute the elements inside, your
  444. 15:32sub-agents, your skills, to have a very
  445. 15:34detailed view of the items. There is
  446. 15:36one per SD in Codex, and on the Cloud
  447. 15:38Code side, you have work areas at the
  448. 15:40sub-agent level. As for skills, here on
  449. 15:43the Codex side, you have the ability to
  450. 15:45create .md skills, same on the Cloud
  451. 15:46Code side. Plugins are directly
  452. 15:49available with MCPs and plugins inside
  453. 15:52Codex. You also have the ability to add
  454. 15:55plugins and MCPs on the Cloud Code side
  455. 15:57. Regarding sub-agents, that is an
  456. 15:59important point, you will be able to
  457. 16:01declare them in Codex, same on the
  458. 16:02Cloud Code side. And we are going to
  459. 16:04create some together right after. You
  460. 16:06can add agents and create your team of
  461. 16:08agents also inside cloud/agent. And on
  462. 16:11the cloud side, you will also be able
  463. 16:13to create sub-agents using/agent. I
  464. 16:15will show you how to write that right
  465. 16:17after. And finally, regarding the state
  466. 16:18, which captures the entire status of
  467. 16:20your loop. You have the ability to
  468. 16:22create .md files in Codex and also in
  469. 16:24Cloud Code to find all the elements and
  470. 16:26progress within them. Ultimately, in
  471. 16:29the end, both tools have the same
  472. 16:31solutions. They just use slightly
  473. 16:32different names each time. And so,
  474. 16:34normally, by now, you have everything
  475. 16:36you need to understand what a loop is,
  476. 16:38what needs to go inside, when to choose
  477. 16:40between an open and a closed loop, the
  478. 16:42elements we include, and how it
  479. 16:44manifests within the solutions you use.
  480. 16:47But I suspect what you really want to
  481. 16:49see is a demo. How can I actually
  482. 16:51create loops in these solutions? What
  483. 16:53does it look like? And is it really
  484. 16:55that powerful? For that, I suggest we
  485. 16:57create a first, very simple loop right
  486. 17:00here inside Claude. This loop. Imagine
  487. 17:03I work for Nike and the idea is that I
  488. 17:05want to track all of Nike's sales. And
  489. 17:07here, to create a first loop, we are
  490. 17:09going to use Claude's Work. Why?
  491. 17:11Because, in fact, Claude's Work can
  492. 17:13already address many elements of a loop
  493. 17:14. Indeed, I can go in and schedule
  494. 17:17tasks, which will allow me to have a
  495. 17:20trigger—the "hit"—that will, for
  496. 17:22example, update items every week. So
  497. 17:25the task will be scheduled, and it can
  498. 17:27run without me. I will also be able to
  499. 17:30connect applications within it, using
  500. 17:31connectors already available here, as
  501. 17:33well as MCP servers. I can attach
  502. 17:37skills to all of this if I need to
  503. 17:39draft items in a certain way. And on
  504. 17:42top of that, inside Claude, I have
  505. 17:44memory. So all elements can be tracked
  506. 17:46and improved. So, in fact, I can
  507. 17:49already create first-level, very simple
  508. 17:51loops using this. For that, here, I'll
  509. 17:54show you a small example. I have a
  510. 17:55folder. This folder is right here. It's
  511. 17:57the Nike sales report. So here I have a
  512. 18:00sales report with various Excel sales
  513. 18:02files that I've had in 2025. Imagine,
  514. 18:04for example, that you had to update
  515. 18:06this file every week. Well, for that,
  516. 18:07you have a PowerPoint here. This
  517. 18:09PowerPoint is right here, and it allows
  518. 18:11for analyzing all of Nike's e-commerce
  519. 18:14performance every week. You have to
  520. 18:16update it every time; it's boring, it's
  521. 18:18tedious, it's a lot of copying. Well,
  522. 18:21for that, a loop here would allow us to
  523. 18:23update it whenever you need it. For
  524. 18:25this, we go back to Cowork, select our
  525. 18:27folder, I choose a folder, and that's
  526. 18:29how I just added the Nike sales report
  527. 18:31here. And what I’m going to ask it to
  528. 18:33do is, every Monday at 10 a.m., update
  529. 18:36the PowerPoint presentation located in
  530. 18:38my folder using the data from the Excel
  531. 18:40files in the 2006 folder. So basically,
  532. 18:43it will open 2006 and take all the new
  533. 18:45data to update my slides. Imagine
  534. 18:48having to update your new Excel files
  535. 18:50every single time they change. You just
  536. 18:53have to drop them in, and then the loop
  537. 18:55will do the work for you. The best
  538. 18:57thing here is using a drive, because
  539. 18:59you don’t even need to upload files,
  540. 19:00and we get a truly 100%autonomous loop.
  541. 19:03Then, once the presentation is updated,
  542. 19:05it will generate the new version of the
  543. 19:07PowerPoint. It will go ahead and write
  544. 19:09an email for me. For example, we could
  545. 19:11have connectors or tools with Gmail
  546. 19:13here. In the body of the email, it
  547. 19:15presents the key results and prepares
  548. 19:17the email so it's ready to be sent. And
  549. 19:19from there, well look, every Monday at
  550. 19:2110 a.m., that’s what will allow for
  551. 19:23the full automation. And from there,
  552. 19:25we're off. And by having it set every
  553. 19:28Monday at 10 a.m., it sees that it's a
  554. 19:30recurring task. So in the end, when
  555. 19:33it’s finished, we’ll have a loop
  556. 19:35running every Monday at 10 a.m. without
  557. 19:37me. It knows where the Excel files are,
  558. 19:40it knows where the PowerPoint is, and
  559. 19:41it knows it needs to send an email. It
  560. 19:44will run continuously, allowing for the
  561. 19:46creation of a first very simple loop
  562. 19:49right here. After a few minutes, if you
  563. 19:51look closely, it clearly understood, as
  564. 19:53we saw, that it is a recurring task. It
  565. 19:55analyzed the folder here and asked me
  566. 19:58for validation. Do I need to have this
  567. 20:00done every week? I gave my approval. It
  568. 20:03did the entire analysis, it has the
  569. 20:05exact structure of all the charts in
  570. 20:06the PowerPoint, and it knows how to
  571. 20:08update them. The script is running
  572. 20:10without errors, it successfully
  573. 20:12performed a test itself to check the
  574. 20:14results, and it has scheduled the task
  575. 20:16to run every Monday. It has analyzed
  576. 20:18the structure and has all the
  577. 20:20components. The next execution is
  578. 20:22scheduled for Monday, July 20th, at 10
  579. 20:24a.m. If I go back here to scheduled
  580. 20:26tasks, you can see that I have my
  581. 20:28weekly PowerPoint and email report
  582. 20:30scheduled every time. So, it could
  583. 20:33update my tasks like this, and I would
  584. 20:35just have to go present them, for
  585. 20:36example, at a steering committee
  586. 20:37meeting. If I want to, I can still run
  587. 20:40the task here once to see the result.
  588. 20:42Here, we can see the task has just
  589. 20:44started, which allows me to run tests
  590. 20:46and ensure the quality of the results.
  591. 20:48And so, as we can see here, it has
  592. 20:50finished updating the entire file with
  593. 20:52the 2026 data. So, it’s basically
  594. 20:54like it updates every Monday; it comes
  595. 20:56in and retrieves the information. And
  596. 20:58if you look here at the top right in
  597. 21:00the output, I have a PowerPoint that
  598. 21:02has been updated. We were on 2025, now
  599. 21:04we are on 2026, and all the numbers
  600. 21:06here have been updated, which would
  601. 21:09allow me, once again, to automate all
  602. 21:11this reporting. And then, as I said, we
  603. 21:14can have it generate a draft email with
  604. 21:16this document as an attachment and thus
  605. 21:19send the email with all the items, all
  606. 21:21automated. I’m just taking a very
  607. 21:24short break in the video, less than a
  608. 21:26minute, but it’s important for those
  609. 21:28who want to go even further in
  610. 21:29mastering AI. Alongside me, you’ll
  611. 21:31find two resources in the description
  612. 21:32of this video that can help you. The
  613. 21:34first is my guide, "The 7 Days of AI."
  614. 21:37It’s a guide I designed to allow you
  615. 21:39to integrate all the basics of AI in
  616. 21:41just 7 days with me. Each day, you
  617. 21:43receive an exclusive resource that will
  618. 21:45help you progress, whether it's
  619. 21:47creating your AI agents, understanding
  620. 21:49the best prompt engineering techniques,
  621. 21:51or getting access to my community of
  622. 21:53over 8,000 members. In short, you
  623. 21:55receive exclusive content by email
  624. 21:56every single day. To get access to it,
  625. 21:58this guide is completely free. You just
  626. 22:01have to come here, enter your email,
  627. 22:02check this box, and press this button.
  628. 22:05And from there, you’re off on the 7
  629. 22:07days of AI, and you’ll receive the
  630. 22:08content for the first day directly. But
  631. 22:11on top of that, I know many of you want
  632. 22:13to automate as many tasks as possible
  633. 22:15using AI. And as you can see in this
  634. 22:17video, I am truly in a position to help
  635. 22:19you. Well, for that, you'll also find
  636. 22:21the presentation link for my AI Master
  637. 22:22masterclass in the description of this
  638. 22:24video. This masterclass is a benchmark
  639. 22:27in France. It's the only masterclass
  640. 22:29that trains you in all the essential
  641. 22:31skills for the AI era. Prompt
  642. 22:33engineering, vibe coding, context
  643. 22:35engineering—you will master all these
  644. 22:37terms and know how to use them. And to
  645. 22:39use them, you must master the best AIs
  646. 22:41on the market. Well, for that, the
  647. 22:43masterclass also trains you on all the
  648. 22:45top AIs available. ChatGPT, Gemini,
  649. 22:47Claude, Perplexity, Microsoft Copilot.
  650. 22:50By the end, you will master all these
  651. 22:52solutions. But above all, right now,
  652. 22:53there is a huge revolution. Agents:
  653. 22:56inside this masterclass, you also have
  654. 22:57an entire module to help you create
  655. 22:59your first AI agents alongside me. You
  656. 23:02can also find testimonials from people
  657. 23:03who have participated in this
  658. 23:05masterclass on the presentation page.
  659. 23:07More than 32,000 people have
  660. 23:08participated since its launch. You
  661. 23:10should know that this masterclass is
  662. 23:11open to everyone, regardless of your
  663. 23:13technical level. Skill building is
  664. 23:15progressive: we learn at your own pace.
  665. 23:17You can pause and come back a bit later
  666. 23:19, just like a Netflix series.
  667. 23:21Personally, however, I commit to one
  668. 23:22thing: AI evolves extremely fast. You
  669. 23:24have undoubtedly noticed it. Well, for
  670. 23:26that, I commit to updating the
  671. 23:28masterclass every month with new
  672. 23:30content. And obviously, from the moment
  673. 23:32you join, you get lifetime access. So
  674. 23:35you benefit from all these updates. So
  675. 23:37don't hesitate to check out those two
  676. 23:38links in the video description. But I
  677. 23:40know you definitely want to go much
  678. 23:42further into loops. Well, for that,
  679. 23:44rest assured, I have planned a much
  680. 23:46more advanced scenario. We will be able
  681. 23:48to create an entire team of AI agents
  682. 23:50that will work in perfect autonomy
  683. 23:52inside a loop. And for that, I chose a
  684. 23:55concrete client case. This client case
  685. 23:57is Trade Republic. And for that, we
  686. 24:00will imagine together that we work in
  687. 24:02the marketing team at Trade Republic
  688. 24:04and want to create a loop of agents
  689. 24:06that works automatically to create our
  690. 24:08Instagram posts, including the
  691. 24:10accompanying text, on the most trending
  692. 24:12topics. For that, what have I built
  693. 24:15here? I built this system that you see
  694. 24:17right here. Inside this system, what is
  695. 24:20it that will allow us to trigger the
  696. 24:22loop? So, having the AI here every
  697. 24:25Monday at 9 a.m., I want a market watch
  698. 24:27on the finance sector to identify the
  699. 24:30most interesting IPO of the week or a
  700. 24:32major market movement, a crisis, an
  701. 24:35opportunity, or a political shift. All
  702. 24:38of these aspects are important to me
  703. 24:40for creating an Instagram post. Because
  704. 24:42if you look here at Trade Republic's
  705. 24:45Instagram page, they share Donald
  706. 24:46Trump's investments, market updates, a
  707. 24:49case study on a specific company like
  708. 24:51Samsung, or the oil market here. They
  709. 24:54cover very specific topics. Here, the
  710. 24:57stock growth of Take-Two, the producer
  711. 25:00of GTA 6, or new IPOs, like SpaceX or
  712. 25:03Anthropic for instance. In short, all
  713. 25:07these elements are what we will try to
  714. 25:10create using this system to have
  715. 25:12LinkedIn posts that look exactly like
  716. 25:14the ones here on specific topics, 100%
  717. 25:17automated. From there, ultimately, the
  718. 25:20loop we're going to create will allow
  719. 25:22us to post on Instagram automatically,
  720. 25:24without me having to deal with it. It
  721. 25:27does this because I have an Instagram
  722. 25:29manager here; it's my planner, my
  723. 25:31orchestrator within the system. It
  724. 25:34distributes tasks to the sub-agents. To
  725. 25:37do this, it triggers every Monday at 9
  726. 25:39a.m. to launch the market watch. It
  727. 25:42calls the Finance Trend Agent each time
  728. 25:44, which is an agent connected to
  729. 25:46various tools. These tools include
  730. 25:49Perplexity. It will go to Perplexity to
  731. 25:52see which finance topic is most
  732. 25:54relevant in France to analyze and
  733. 25:56communicate about this week. So, for
  734. 25:59this, my orchestrator launches the
  735. 26:01agent. In the end, it has identified a
  736. 26:02topic. It then gives this topic to the
  737. 26:05copywriter. The copywriter is the
  738. 26:07expert in writing the Instagram caption
  739. 26:09with the right keywords, perfectly
  740. 26:11aligned with the chosen topic. And
  741. 26:14obviously, it replicates the exact tone
  742. 26:15and style of the posts we see on the
  743. 26:17Trade Republic Instagram. How did I do
  744. 26:20that? Well, you'll see, I'm going to
  745. 26:22train a skill called "Text DNA" that
  746. 26:24understands, based on all of Trade
  747. 26:25Republic's Instagram posts, how they
  748. 26:27are written. And finally, I have one
  749. 26:30last sub-agent specialized solely in
  750. 26:32image creation. For that, it has also
  751. 26:35been trained. It was trained on the
  752. 26:37Instagram posts that I scraped directly
  753. 26:39from the Trade Republic account. And to
  754. 26:42create the image, it has a tool, it has
  755. 26:44a nano banana that allows it to
  756. 26:45understand how it should create and
  757. 26:47have a good image generation tool. Once
  758. 26:50they have that, they are able to
  759. 26:52generate all the elements for me, they
  760. 26:53send them back to the Instagram manager
  761. 26:55who presents the result to me.
  762. 26:57Obviously, here, I chose to have the
  763. 26:59result presented to me. If we wanted to
  764. 27:02go even further, I could connect
  765. 27:03another tool that would allow me to
  766. 27:05post this directly to Instagram
  767. 27:07automatically, but here, I preferred to
  768. 27:09stop to be able to see the result with
  769. 27:11you before publishing it. And so, what
  770. 27:14we are going to do now is go and create
  771. 27:16this whole system together, and
  772. 27:18obviously, you can find all the prompts
  773. 27:20to create this yourself in my community
  774. 27:22using my free guide in the video
  775. 27:23description, the 7 days of AI. So, in
  776. 27:26order to create this whole loop system
  777. 27:28for Trade Republic here, automating the
  778. 27:30publication of my posts on Instagram, I
  779. 27:32already had to do several things. Right
  780. 27:34there, if we look closely, I created a
  781. 27:36folder called Trade Republic Marketing.
  782. 27:38And inside, if we look closely here, I
  783. 27:41retrieved some Instagram posts from
  784. 27:43Trade Republic. I took screenshots of
  785. 27:45them and we can see the post
  786. 27:46description on the right here and the
  787. 27:48type of image we want to reproduce on
  788. 27:49the left. And so, we are just going to
  789. 27:51need that to initiate the loop. And so
  790. 27:54now to initiate all of this, we are in
  791. 27:55Cloud Code. In Cloud Code here, you are
  792. 27:57going to select a project. For me, this
  793. 27:59project is very simple here, it's Trade
  794. 28:01Republic marketing. The first thing we
  795. 28:03are going to do is start by initiating
  796. 28:05the entire context of this project. For
  797. 28:06that, right here, I am going to send a
  798. 28:08first prompt. This prompt will allow me
  799. 28:10to provide all the instructions and the
  800. 28:12general context. I am the Trade
  801. 28:13Republic marketing team. The goal here
  802. 28:16of the agent system is to produce
  803. 28:18high-performing Instagram posts for
  804. 28:19Trade Republic. The structure, so I
  805. 28:22explain all the agents that I am going
  806. 28:24to put inside my structure. So, it's
  807. 28:26the little diagram that I presented to
  808. 28:28you previously. I explain to it that
  809. 28:29there will be a manager agent and then
  810. 28:31sub-agents. I also give it reference
  811. 28:34sources, the website, the Trade
  812. 28:36Republic Instagram account, the
  813. 28:37editorial objectives, the fundamental
  814. 28:39rules, which will allow me to provide
  815. 28:41the entire execution framework for my
  816. 28:43agent system. And here, as you can see,
  817. 28:46it has created this little .md file. So
  818. 28:48this allows you to start using it in
  819. 28:50Claude Code right here. But personally,
  820. 28:52I prefer to use it directly in Cursor.
  821. 28:55For that, you just come right here and
  822. 28:57select open in Cursor. From there,
  823. 28:59Cursor will open and you'll see we can
  824. 29:00use Claude Code without being a
  825. 29:02developer to follow through the
  826. 29:03scenario. If you look here, we already
  827. 29:06find what it just generated for me in
  828. 29:08Claude Code, and we can see the
  829. 29:10Instagram posts here, for example, with
  830. 29:12the different images since we selected
  831. 29:14the Trade Republic marketing folder.
  832. 29:17Now, we can ask Claude Code on the
  833. 29:20right to use all the Instagram posts in
  834. 29:23the "Instagram posts" folder, and also
  835. 29:26consult the .mmd file we just created
  836. 29:29with all the context, to generate a
  837. 29:32skill capable of understanding the tone
  838. 29:35, format, and atmosphere of the images
  839. 29:38—and only the images. Why? Because it
  840. 29:43will allow me to design a skill that
  841. 29:45can then tell subagents how images
  842. 29:47should be created. So, you launch
  843. 29:50Claude Code and it will start designing
  844. 29:52the skill for you. As you can see now,
  845. 29:54it has finished. The skill has been
  846. 29:56generated here. It’s called Instagram
  847. 29:59Visual DNA. Now, what I suggest we do
  848. 30:02is the same thing, but focusing only on
  849. 30:04the description text that accompanies
  850. 30:05each of the images in my Instagram
  851. 30:07posts. We send the request to generate
  852. 30:10the skill. Here again, the skill
  853. 30:12creation is finished. We now have
  854. 30:14Instagram Text DNA, which is a skill
  855. 30:17that allows me to generate this text.
  856. 30:20Now that we have that, if we look here
  857. 30:22and type/mcp server, we can see that
  858. 30:25several MCP servers are already
  859. 30:27connected. Indeed, I have already
  860. 30:30connected the MCP server with
  861. 30:31Perplexity, because if we go back to
  862. 30:33the copywriter system, we just created
  863. 30:35this skill. We also created the skill
  864. 30:38here for the visuals. Now, we will
  865. 30:41start creating the agent here for
  866. 30:43researching interesting finance topics.
  867. 30:46And for that, it needs to rely on
  868. 30:47Perplexity here. So, type/agent here,
  869. 30:50which will open the terminal, and from
  870. 30:52there, you will be able to create a new
  871. 30:54agent for a project you’re generating
  872. 30:56with Claude. You paste the information
  873. 30:59into it. Here, we want to create an
  874. 31:01agent specialized in researching
  875. 31:03current finance topics. What kind of
  876. 31:05finance topics are we talking about
  877. 31:07here? IPOs, growing or shrinking
  878. 31:10industries, key sectors, major
  879. 31:12political shifts; these are all
  880. 31:13important topics for us. Then, to
  881. 31:18conduct research, you must use the
  882. 31:20Perplexity MCP server, which will
  883. 31:22automatically search for information
  884. 31:25from the last three days to ensure the
  885. 31:27selected topics are current and
  886. 31:30relevant. We send the info, and the
  887. 31:32agent is being generated. Here, the
  888. 31:34agent is asking me which tools it can
  889. 31:36access. We’re going to give it all of
  890. 31:37them. It will run with a name, and
  891. 31:39we'll give it a little blue color to
  892. 31:41remind us of Perplexity. We link it to
  893. 31:44the project for memory. And here we go.
  894. 31:46The agent is created. As you can see
  895. 31:47here, Finance News Scout is the agent
  896. 31:49that was created. Now, let's ask it to
  897. 31:51create a second agent. This time, the
  898. 31:54goal is to tell it to create an agent
  899. 31:56specialized in creating Instagram post
  900. 31:58images. Its mission is to produce the
  901. 32:01visual, and only the visual. To do this
  902. 32:04, it will take the current news topic
  903. 32:06from our Finance agent, which will have
  904. 32:08analyzed the info in Perplexity. You
  905. 32:10never identify the topics to cover
  906. 32:12yourself. Next, I gave this sub-agent
  907. 32:14several steps to follow, but most
  908. 32:16importantly, it must respect the visual
  909. 32:18identity. For that, it uses the
  910. 32:21Instagram visual DNA skill we created
  911. 32:23previously. And to create the image, it
  912. 32:26automatically relies on the Nano Banana
  913. 32:28MCP. to create the image. I gave it a
  914. 32:31few constraints. It must only create an
  915. 32:33image and should never write text,
  916. 32:35draft a post, suggest hashtags, and so
  917. 32:38on. Its goal is strictly to make the
  918. 32:40image. I chose this deliberately. This
  919. 32:42way, I ensure that the final result is
  920. 32:44of very high quality. Now, as we can
  921. 32:46see, it has successfully created the
  922. 32:47agent. We can see the Finance News
  923. 32:50Scout agent here, and the Instagram
  924. 32:53Visual Creator here. So, I have my two
  925. 32:56sub-agents ready. I will need to create
  926. 32:58one last one. This one follows more or
  927. 33:00less the same process as the one we
  928. 33:02just created for the images. However,
  929. 33:04it will be focused solely on the text.
  930. 33:06This way, I have my sub-agent to find
  931. 33:08the relevant topic. And I have my two
  932. 33:10agents to create the image as well as
  933. 33:12the text. And once we've done that,
  934. 33:14only one step will remain, which is to
  935. 33:17create the planner, the orchestrator
  936. 33:19for these sub-agents, right after. And
  937. 33:21so now, I have my three sub-agents
  938. 33:23ready to go. If we look here for a
  939. 33:25second, we have this sub-agent that's
  940. 33:27ready. We have the other two here as
  941. 33:28well. They have access to their tools
  942. 33:31and various skills to ensure the
  943. 33:32consistency of the creations. Now, we
  944. 33:35are going to create the Instagram
  945. 33:36manager that will oversee all of this.
  946. 33:38For that, I come to the right, and here
  947. 33:40, if you look closely, you are the
  948. 33:41manager of my project's agents. We
  949. 33:43explain his entire mission to him. We
  950. 33:45tell him that, above all else, he must
  951. 33:46systematically check the folder. And
  952. 33:48here, we'll put trade republic
  953. 33:50Instagram agent.md with all the context
  954. 33:52. The goal is to create Instagram posts
  955. 33:54for Trade Republic. I've put the entire
  956. 33:57process here for him. First, he
  957. 33:58launches the Finance News Scout agent
  958. 34:01to go find relevant topics. Next, he
  959. 34:04creates the visual, and then he creates
  960. 34:05the text. At the end, he provides me
  961. 34:08with a full summary, the selected
  962. 34:09topics, visual concepts, the caption,
  963. 34:11and an explanation. I'm launching the
  964. 34:13creation of this last agent. Now, it's
  965. 34:15finished. We can see that it has
  966. 34:16created all the different agents for me
  967. 34:17. Now, what interests us is seeing if
  968. 34:20they are capable of working together
  969. 34:21and providing me with a good result.
  970. 34:24For that, I am starting a new session
  971. 34:25here. I ask the manager agent right
  972. 34:28here, the Instagram manager, and I tell
  973. 34:30him to start the process for the next
  974. 34:33Instagram post. And here we go. You can
  975. 34:35see that the full cycle has started.
  976. 34:38The first thing it just did, as I
  977. 34:39requested, is read all the context in
  978. 34:41the md file. And then here, it's
  979. 34:44calling the Scout agent to go perform
  980. 34:46the analysis on Perplexity and pull out
  981. 34:49an interesting topic for an Instagram
  982. 34:51post. You can see, by the way, right
  983. 34:54here, I can see the whole cycle, all
  984. 34:56the steps that will need to be
  985. 34:57triggered by the manager agent until
  986. 34:59reaching step 4b to output the entire
  987. 35:01Instagram post. And now here, after a
  988. 35:04few minutes, we see that the agent has
  989. 35:06completed all the tasks. If you look
  990. 35:08here, it has already tracked its entire
  991. 35:10to-do list. It went to retrieve the
  992. 35:12information here using the finance
  993. 35:15agent. It has brought back a topic.
  994. 35:18Then, it proceeded to use copywriting
  995. 35:20to write the post and the creative
  996. 35:21agent to generate the image. The topic
  997. 35:24it chose was the IPO of the brand Le
  998. 35:27Slip Français. So, this is going to be
  999. 35:29extremely interesting. Plus, it's
  1000. 35:30French, so that’s really cool. And
  1001. 35:32we're going to take a look at the
  1002. 35:33result it generated here. In the end,
  1003. 35:35if you look closely, it made all the
  1004. 35:37edits and approved an Instagram post
  1005. 35:39for the next cycle. Here, the Le Slip
  1006. 35:42Français IPO is taking place on July
  1007. 35:4414th; at the time I'm filming this
  1008. 35:46video, it's the 17th, so we’re well
  1009. 35:48within the three-day window that was
  1010. 35:51set to create a small carousel. It
  1011. 35:53prepared the first image for me, and
  1012. 35:55then it prepared five prompts for the
  1013. 35:57next visuals if I need them. It
  1014. 35:59provided the caption with the text to
  1015. 36:01be filled in. It also included a
  1016. 36:03virality potential score. Now, we’re
  1017. 36:05going to see what this image looks like
  1018. 36:07, because we already have the text
  1019. 36:09right here. Here is the image. It
  1020. 36:11perfectly respects what Le Slip
  1021. 36:13Français does. We have the blue, white
  1022. 36:15, and red IPO flag here; Le Slip
  1023. 36:17Français has just gone public. Here is
  1024. 36:19how. And here, it really captures and
  1025. 36:22reflects the DNA and the structure of
  1026. 36:23the posts we have on Trade Republic. If
  1027. 36:26you look closely, we have the large
  1028. 36:27text at the top with a bit of text
  1029. 36:29right below it every time. Here, it
  1030. 36:31provided it in French for us. There you
  1031. 36:33go, we have the same patterns every
  1032. 36:34time. Well, not on these posts here,
  1033. 36:36but here for example, we have all the
  1034. 36:38right elements every time. The text is
  1035. 36:41ready to go as well. It would just need
  1036. 36:43to be published. Now, the last thing
  1037. 36:46left to do is to reopen Cloud Code here
  1038. 36:48and tell it: "Make sure this folder and
  1039. 36:50the agents running in it act as a loop
  1040. 36:53to deliver the post every Monday at 9
  1041. 36:55a.m." And from there, now that I've
  1042. 36:57tested the loop and it works, obviously
  1043. 36:59, you could still make changes if
  1044. 37:01needed. That's this testing phase. But
  1045. 37:04here, let's imagine I want to continue;
  1046. 37:06well, I can set up the loop. And from
  1047. 37:09that point on, I don't have to do any
  1048. 37:10more work at all. This agent system
  1049. 37:13will run directly in a 100%autonomous
  1050. 37:15way. And there it is, right here. After
  1051. 37:18a few minutes, it’s all set up. Here
  1052. 37:19is what was created. He created all the
  1053. 37:21automation. Here, if you look closely
  1054. 37:24on the left, I now have automation with
  1055. 37:26the weekly prompt here, which lets me
  1056. 37:28start the loop without me. I don't have
  1057. 37:31to type prompts anymore. And here, we
  1058. 37:33have the entire loop script that is now
  1059. 37:35able to run without me. And so here,
  1060. 37:39well, we managed to create a 100%
  1061. 37:40automated loop together in just a few
  1062. 37:42minutes, which would allow me to
  1063. 37:44generate Instagram posts. Once again,
  1064. 37:47if you want to go a little bit further
  1065. 37:49here, you add a tool capable of posting
  1066. 37:51directly to Instagram for you or doing
  1067. 37:54it via computer use, as is already
  1068. 37:56possible with Claude, and from there
  1069. 37:58you are on a process. using the loops.
  1070. 38:02So, normally with this video, you have
  1071. 38:04everything you need to create your
  1072. 38:06first loops and automate even more
  1073. 38:07tasks with AI. And as you saw in the
  1074. 38:10video, inside the loop, we created a
  1075. 38:11small team of AI agents. But perhaps
  1076. 38:14you want to create a much more complete
  1077. 38:16team of agents using Claude. Well,
  1078. 38:18that's perfect because right here you
  1079. 38:20have a video displayed where I show you
  1080. 38:22how to create your first team of agents
  1081. 38:24using Claude.

About this transcript

This page contains the full transcript of Arrêtez les prompts. Passez aux Loops (Le futur des agents IA) by Elliott Pierret, generated from the public captions YouTube serves with the video. The transcript has 7,430 words across 1,081 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.