Arrêtez les prompts. Passez aux Loops (Le futur des agents IA) — Transcript
Full transcript
- 0:00You need to stop creating agents, and
- 0:02it’s not just me saying it; it’s
- 0:04Boris Chouhy, the creator of Claude
- 0:06Code, and Peter Steinberger, the
- 0:08creator of OpenClow, saying it. And for
- 0:10them, they summarized it in one
- 0:11sentence. They no longer write prompts
- 0:13in Claude. Instead, they write loops
- 0:16that prompt for them in Claude and
- 0:18decide which actions to take. Basically
- 0:21, their new job is to write loops. But
- 0:23the problem is that the word "loop"
- 0:24doesn't mean anything to most people.
- 0:26What is a loop used for? What does it
- 0:28look like? When should we create one,
- 0:30and most importantly, how is it
- 0:31actually built? Well, in this video,
- 0:33I’ll clear all of that up. I will
- 0:35define what a loop is and what you need
- 0:36to put inside it. I will also show you
- 0:39many use cases for loops, but more
- 0:41importantly, I’ll provide concrete
- 0:43demonstrations for Nike and Trade
- 0:44Republic, where I’ll create loops
- 0:46that will help you understand the true
- 0:48value of this new technique that will
- 0:50totally change the way you work with AI
- 0:52. Let’s get started creating your
- 0:55first loops. So, first, to really
- 0:56understand this concept of a loop, we
- 0:58need to look back together at the AI
- 1:00generations we’ve used so far. If you
- 1:03look closely, as we’ve done on the
- 1:05channel, we started by trying to
- 1:07generate better answers using the AIs
- 1:09we use. At the beginning, the model
- 1:12produced text based on a prompt, and in
- 1:14the end, we had an answer, and all the
- 1:16value resided in the quality of that
- 1:18response. And you, well, you were
- 1:21responsible for writing the right
- 1:23prompts to increase the quality of that
- 1:25response, and then came the age of
- 1:27agents. And there, the goal was no
- 1:30longer just to write good prompts, but
- 1:32to create systems capable of acting
- 1:34perfectly based on a request. Then our
- 1:37agent would call tools, it had memory,
- 1:39and in the end, it was capable of
- 1:41performing actions for us. We set goals
- 1:43, provided a work environment, and in
- 1:45the end, we had a result. But now, with
- 1:48this notion of a loop, we’re arriving
- 1:50at a whole new generation of agents.
- 1:52The idea is to build systems capable of
- 1:54learning on their own. Ultimately, we
- 1:57give it a framework within which it can
- 1:59execute. Then, it performs tasks. It
- 2:02then evaluates the result itself or
- 2:04with our help. It will then correct
- 2:07itself, and in the end, it will
- 2:08understand the adopted behaviors so
- 2:09well that it will be able to do it
- 2:11alone without our intervention. Over
- 2:13the course of executions, it will
- 2:15improve, and the result will always be
- 2:17of the same or better quality. If you
- 2:19look, you can see it even better here.
- 2:21At first, we typed prompts, called the
- 2:23model, and got the answer. Now, we have
- 2:25a system that will be able to produce
- 2:28actions. What kind of actions? Well,
- 2:30here, the actions we will have defined
- 2:32ourselves under our supervision where
- 2:34we gave the trajectory, it will produce
- 2:36actions. Then, it will evaluate the gap
- 2:39between what it provided and what we
- 2:41initially wanted. And, in the end, it
- 2:44will correct itself to ensure that the
- 2:46next time it produces an action, it
- 2:49will be of good quality, always with
- 2:51supervision from us, but also its own
- 2:53supervision and control within the
- 2:55system. So what you need to remember
- 2:59here is that when we create loops in
- 3:01this new generation of AI, the goal
- 3:02will no longer be to provide a prompt
- 3:04and get a result, but to explain
- 3:06everything that happens inside a system
- 3:08. And this system, in the end, must be
- 3:11able to run without you, and you just
- 3:13have to watch it, supervise it, and
- 3:14then come in to rectify it if there are
- 3:16occasional discrepancies. But otherwise
- 3:19, the system is self-learning,
- 3:21self-correcting, and always provides
- 3:23you with the results you need at
- 3:24certain times, under certain conditions
- 3:26. And so, once we've said that, I
- 3:28suggest we now go back to the loop
- 3:30architecture. What do we need to put
- 3:32into these loops for them to work? If
- 3:35you look here, it comes from an
- 3:37analysis done by the Anthropic teams,
- 3:39who share the best possible
- 3:40architecture for what we call a
- 3:42recursive loop. One that is capable of
- 3:45never stopping and having a continuous
- 3:47cycle. We will have a planner that
- 3:49breaks the objective down into a
- 3:51realistic plan. It ultimately provides
- 3:54the plan and the steps to accomplish,
- 3:55and above all, it sets the course with
- 3:57the final goal. Next, we have the
- 4:00builder. It applies the plan and calls
- 4:03the right tools at the right time. You
- 4:05need to send an email, it calls Gmail,
- 4:07view data in the CRM, or retrieve a
- 4:09customer record from your ERP. It is
- 4:12the one that implements the plan and
- 4:13calls the right tools. Next, well,
- 4:15within this plan, there may be some
- 4:16creation, some elements to be done.
- 4:18Ultimately, the bulk of the work is
- 4:20done by the builder. Then, once it's
- 4:23finished, we'll put a verifier in the
- 4:25loop to ensure that the output meets
- 4:27our initial expectations. Once we have
- 4:30that, there's a super important role in
- 4:32agents since their inception, which is
- 4:34memory. How do we ensure that within
- 4:37the system, we have a memory that can
- 4:39understand successes, but also failures
- 4:42, and thus self-improve the elements
- 4:44within the loop. Next, we'll see the
- 4:46manager who will ultimately review the
- 4:48plan and see if everything has been
- 4:50done or not. And at the end, we have
- 4:52the controller. The controller is
- 4:53extremely important because it reviews
- 4:55the entire loop method. It will go into
- 4:57the memory to see, ultimately, were
- 4:59there failures, were there successes?
- 5:01If so, were there missing elements that
- 5:03led to a failure? Were there elements
- 5:06we added that were extremely relevant
- 5:08and led to a success? All of this
- 5:11consists of elements inside the loop.
- 5:13And when the controller is finished,
- 5:14the goal is not for the loop to stop.
- 5:17The controller takes all this
- 5:19information, gives it to the planner,
- 5:21who, when breaking down the objective
- 5:23into a future plan with actions to
- 5:25perform, will have taken all of this
- 5:27into account, and normally, the chain
- 5:29is self-improved, so that in the end,
- 5:31we can ensure we always have the best
- 5:33result possible. If you look here, I've
- 5:36provided a more detailed view of these
- 5:38different steps where the planner
- 5:40breaks down the objective into a
- 5:41realistic plan. The builder applies the
- 5:44plan, calls the tools, and produces the
- 5:46results. The verifier checks what has
- 5:48been produced. Memory keeps track of
- 5:50everything that worked or didn't work.
- 5:53The manager classifies what remains to
- 5:55be done and redirects the subsequent
- 5:57operations. And the controller reviews
- 5:59the entire method then passes the baton
- 6:02to the planner to rerun the whole loop.
- 6:04And so, all this is the theoretical
- 6:06part of building loops. But you're
- 6:08probably wondering how all this
- 6:10materializes. Well, right here I've
- 6:12laid out the building blocks of a
- 6:14functional loop. If you look at this
- 6:16loop, it picks up a project from
- 6:18Encloud here. We could have a loop as
- 6:21shown here, and inside this loop, there
- 6:23are key elements that need to be found
- 6:25every time. And regardless of whether
- 6:28you work in the Cloud or at OpenAI, we
- 6:30find the same elements. The first
- 6:32concept is the notion of a habit.
- 6:34Without it, the loop won't start, and
- 6:36you would need human intervention. What
- 6:39we want is for it to run automatically.
- 6:41What can trigger a loop can be a
- 6:43schedule. Every Monday at 9 a.m., it
- 6:46runs, or once a month, or once a year.
- 6:49For example, every Monday at 9 a.m. is
- 6:51what would trigger a loop to perform
- 6:53reporting actions. Another example is a
- 6:56goal. A goal is an objective we set
- 7:00within the loop that continues to run
- 7:02until it succeeds. In that case, we are
- 7:05looking at long-term projects. Then,
- 7:07you have the hooks. Hooks are what
- 7:10trigger commands at specific moments.
- 7:12For instance, a follow-up cycle could
- 7:15trigger a loop. If you want to follow
- 7:17up with your clients every 7 days or
- 7:19when you haven't received a reply after
- 7:21a certain time, the loop triggers to
- 7:23contact the client. Imagine this for
- 7:25customer support cases as well.
- 7:27Whenever you receive a new support
- 7:29ticket, the loop triggers from a
- 7:31solution and performs all the tasks for
- 7:34you. Another loop system could be that
- 7:37every time you create a quote and save
- 7:39it to your drive, it triggers reporting
- 7:41actions, for example. These are all
- 7:44technical elements, but they send a
- 7:46signal to trigger the loop and perform
- 7:48various actions. Here, the loop has
- 7:50been triggered. Next, we find things
- 7:53like skills inside. This is a concept
- 7:55you should normally be familiar with in
- 7:57AI. Skills are ultimately how you set
- 8:00rules that, once written, can be
- 8:02replayed by the AI. I always want you
- 8:05to write follow-up emails in this
- 8:06format. If you need to generate a quote
- 8:09automatically, here is how our quotes
- 8:11are drafted. If you are doing reporting
- 8:13, here is the slide template to adopt.
- 8:16Ultimately, these are all the elements
- 8:18within the loop that define how work is
- 8:20performed inside the loop. And for that
- 8:23, we will rely on the skills we have
- 8:25trained beforehand. Next, inside the
- 8:27loop, there will be one or more agents.
- 8:30Generally, as you'll see, we have a
- 8:32central agent at the top who then
- 8:34delegates tasks to sub-agents. The idea
- 8:37is that we divide the tasks and
- 8:38therefore the work among the agents.
- 8:41There will be the one who produces, the
- 8:42one who validates, the one who plans.
- 8:45This allows us to avoid hallucinations
- 8:47and ensure the work is well distributed
- 8:49, so each agent can call various skills
- 8:51and connectors. Speaking of connectors,
- 8:55you will obviously put tools and
- 8:57connections inside your loop. Tools
- 9:00that will allow you to broaden the
- 9:03context and the scope of your loop. You
- 9:05want to send an email, well, you'll add
- 9:07the Gmail connector. You want to be
- 9:09able to see your customer data in the
- 9:10CRM, you'll add Salesforce. You will
- 9:12add all the connectors or MCP servers
- 9:15that will come inside the loop at the
- 9:17level of the different agents to call
- 9:19your context and execute a specific
- 9:21task using skills. And finally, as I
- 9:24mentioned previously, there is this
- 9:26notion of memory which is extremely
- 9:28important. Why? Because as the loop
- 9:31runs, well, the memory will expand. The
- 9:35memory will allow you to specify how
- 9:37you want certain things, but above all,
- 9:39the memory will recover everything that
- 9:41happened in the loop each time, to
- 9:43ensure that, for example, if a
- 9:45sub-agent produced erroneous results
- 9:47two or three times, it can self-modify
- 9:49following an inspection by another
- 9:51agent to improve performance, without
- 9:53you or me having to intervene in the
- 9:55loop to manually modify one of the
- 9:57agents. The agents will talk to each
- 10:00other. One will see that another
- 10:02performed poorly, and as time goes on,
- 10:05with the help of memory, we will
- 10:07understand why and who failed in order
- 10:10to self-improve. So now you should
- 10:12normally understand what a loop is and
- 10:14the elements we need to put inside it.
- 10:15But then there is a fundamental
- 10:17question. Should I leave the loop open
- 10:20or should I close it? And for that,
- 10:22right here, I have given you two
- 10:23examples. Open loops are loops that I
- 10:25recommend you use for scenarios, for
- 10:28example, aimed at the general public.
- 10:30You'll put them on your websites with
- 10:32an agent that responds to emails or an
- 10:34agent that processes incoming requests
- 10:35on a landing page. The idea is that the
- 10:38agent will have a very, very wide
- 10:39search space. The process has so many
- 10:42possibilities that we need to have the
- 10:44widest possible loop to handle requests
- 10:46. The agent will wander around inside
- 10:48and try to find the best possible
- 10:50result. However, expect a high
- 10:52consumption of tokens at that point.
- 10:55The idea here is more for big scenarios
- 10:56where the ROI is so strong that the
- 10:58budget is almost unlimited. So, we can
- 11:00afford to have this learning phase at
- 11:02the beginning. Be careful, just because
- 11:04the loop is open and very wide doesn't
- 11:06mean it will always stay that way.
- 11:08Ultimately, the idea of the loop, once
- 11:10again, is that it should be
- 11:12self-learning, so after a while, we can
- 11:14imagine that inside the loop, the field
- 11:16of possibilities narrows to provide
- 11:18much higher-quality answers. The
- 11:20concept is fewer barriers but more
- 11:22creativity. The closed loop is really
- 11:24what I recommend you start with. The
- 11:27idea is that it will be a closed
- 11:28process where you provide the different
- 11:30steps to be completed. That is exactly
- 11:33what we will create together in the
- 11:34demo shortly. The idea inside is that
- 11:37we will have a clear evaluation of each
- 11:39step. We will have a more restricted
- 11:41budget because it will consume fewer
- 11:43tokens. But most importantly, your loop
- 11:45is very structured. You know the final
- 11:47result you are looking for, so
- 11:49naturally, you will create the loop
- 11:50accordingly, starting from what you
- 11:52want to achieve and including all the
- 11:54necessary elements in the loop every
- 11:55time. So, at the beginning, stick more
- 11:57to closed loops, but know that you can
- 11:59also move on to open loops depending on
- 12:01your scenarios. And then you might ask
- 12:03me, "OK great, but at what point should
- 12:05I actually create a loop?""Why can't I
- 12:07just keep using an agent or just chat
- 12:09with the AI?" Well, for that, I’ve
- 12:11created a little checklist I'm going to
- 12:13share with you, which will help you
- 12:15understand if your scenario works with
- 12:17a loop or not at all. For that, right
- 12:19here, you have four conditions that I
- 12:21recommend you apply to know if a loop
- 12:24is relevant or not. The first condition
- 12:26is, does your task repeat itself
- 12:27multiple times? Because you need to do
- 12:30it every day or every week. Based on
- 12:32that, it is necessarily considered
- 12:34recurring. So from there, it is a good
- 12:36condition to start creating loops. Next
- 12:38, the second condition is, is there a
- 12:40possibility of failure? Here, the idea
- 12:43is: are you able to clearly identify
- 12:45when it worked and when it didn't work?
- 12:48To do that, if you're not able to tell
- 12:50yourself "this is a failure, this is a
- 12:52success" every time in a very binary
- 12:54way, then there’s really no point in
- 12:56using a loop. Why? Because in the end,
- 12:59you will always have to go and verify
- 13:00the final result. So, there’s an
- 13:02intervention on your part. You aren't
- 13:04fully in supervision and delegation
- 13:06mode. No, you have to come and see all
- 13:08the results every time and check them.
- 13:11That’s not the goal of a loop at all.
- 13:12So, if that’s your situation,
- 13:13definitely go with a different scenario
- 13:15. However, if you are able to identify
- 13:17"this works, that doesn't," and the
- 13:19loop can do it for you without
- 13:21supervision, then you have a scenario
- 13:23for a loop. Next, can the loop do all
- 13:25the work, or do you have to come in at
- 13:27some point to press a button? Uh, do
- 13:29you have to enter text manually at some
- 13:31point because you're the one who has
- 13:32the information in your head, for
- 13:34example? Well here, it's super
- 13:35important; the goal is for the entire
- 13:37loop to function end-to-end without you
- 13:39, and in that case, you have another
- 13:41condition met to create your loop. And
- 13:44finally, the last condition is: are you
- 13:46able to say, "Okay, this is finished,
- 13:48there is no human interpretation or
- 13:50judgment in what was done"? And this is
- 13:53certainly where you will have the most
- 13:54difficulty identifying a loop. But,
- 13:56you’ll see that we still arrive at
- 13:57things where, after a while, we’ve
- 13:59run so many tests. It’s on this phase
- 14:01, by the way, that the testing part is
- 14:03super important. You’ll test the
- 14:05results 10, 15, 20 times. After a while
- 14:08, if at the end of these 15 or 20
- 14:09results you aren't able to say, "It's
- 14:11good, all 20 are correct."...then from
- 14:13that point on, it means there is human
- 14:14judgment involved. There is too much
- 14:16interpretation for it to be done by a
- 14:18loop, and a human will always have to
- 14:19verify it at the end. And in that case,
- 14:21it’s not necessarily the goal of a
- 14:22loop. If you tell yourself, "Okay, it
- 14:24doesn't need me anymore." You have the
- 14:25conditions and you can create your loop
- 14:27. So now, you have everything you need
- 14:28to know when to create a loop. But to
- 14:30create one, you need the right tools.
- 14:32For this today, I have chosen two tools
- 14:34: Codex and Cloud Code. In my opinion,
- 14:36they are the two best tools today for
- 14:37creating loops. Later, in the demo
- 14:39section, we will create loops inside
- 14:41Cloud with Cowork as well as with Cloud
- 14:43Code. But I still wanted to go over the
- 14:46elements of a loop with you. Ultimately
- 14:48, what do they look like inside these
- 14:49solutions? What building blocks will
- 14:51you need to use in these solutions to
- 14:53create your loop? And for that, I’ve
- 14:55made a small table here with the
- 14:56different elements for automation. So,
- 14:59what will allow you to trigger, for
- 15:01example, the Heartbeat aspect I
- 15:02mentioned earlier? Well, for that, in
- 15:04Codex you have a section called
- 15:06automation which is available in the
- 15:08app. You just have to go there. You
- 15:10will be able to create small loops
- 15:11inside and set goals. Otherwise, on the
- 15:14Cloud Code side, as we will see, you
- 15:17can use/loop or/g to launch the various
- 15:19elements. There is also the whole hook
- 15:22part that will allow you to trigger
- 15:24from a behavior in an application, for
- 15:26example. Next, you will have the work
- 15:28areas. This is what will allow you to
- 15:30distribute the elements inside, your
- 15:32sub-agents, your skills, to have a very
- 15:34detailed view of the items. There is
- 15:36one per SD in Codex, and on the Cloud
- 15:38Code side, you have work areas at the
- 15:40sub-agent level. As for skills, here on
- 15:43the Codex side, you have the ability to
- 15:45create .md skills, same on the Cloud
- 15:46Code side. Plugins are directly
- 15:49available with MCPs and plugins inside
- 15:52Codex. You also have the ability to add
- 15:55plugins and MCPs on the Cloud Code side
- 15:57. Regarding sub-agents, that is an
- 15:59important point, you will be able to
- 16:01declare them in Codex, same on the
- 16:02Cloud Code side. And we are going to
- 16:04create some together right after. You
- 16:06can add agents and create your team of
- 16:08agents also inside cloud/agent. And on
- 16:11the cloud side, you will also be able
- 16:13to create sub-agents using/agent. I
- 16:15will show you how to write that right
- 16:17after. And finally, regarding the state
- 16:18, which captures the entire status of
- 16:20your loop. You have the ability to
- 16:22create .md files in Codex and also in
- 16:24Cloud Code to find all the elements and
- 16:26progress within them. Ultimately, in
- 16:29the end, both tools have the same
- 16:31solutions. They just use slightly
- 16:32different names each time. And so,
- 16:34normally, by now, you have everything
- 16:36you need to understand what a loop is,
- 16:38what needs to go inside, when to choose
- 16:40between an open and a closed loop, the
- 16:42elements we include, and how it
- 16:44manifests within the solutions you use.
- 16:47But I suspect what you really want to
- 16:49see is a demo. How can I actually
- 16:51create loops in these solutions? What
- 16:53does it look like? And is it really
- 16:55that powerful? For that, I suggest we
- 16:57create a first, very simple loop right
- 17:00here inside Claude. This loop. Imagine
- 17:03I work for Nike and the idea is that I
- 17:05want to track all of Nike's sales. And
- 17:07here, to create a first loop, we are
- 17:09going to use Claude's Work. Why?
- 17:11Because, in fact, Claude's Work can
- 17:13already address many elements of a loop
- 17:14. Indeed, I can go in and schedule
- 17:17tasks, which will allow me to have a
- 17:20trigger—the "hit"—that will, for
- 17:22example, update items every week. So
- 17:25the task will be scheduled, and it can
- 17:27run without me. I will also be able to
- 17:30connect applications within it, using
- 17:31connectors already available here, as
- 17:33well as MCP servers. I can attach
- 17:37skills to all of this if I need to
- 17:39draft items in a certain way. And on
- 17:42top of that, inside Claude, I have
- 17:44memory. So all elements can be tracked
- 17:46and improved. So, in fact, I can
- 17:49already create first-level, very simple
- 17:51loops using this. For that, here, I'll
- 17:54show you a small example. I have a
- 17:55folder. This folder is right here. It's
- 17:57the Nike sales report. So here I have a
- 18:00sales report with various Excel sales
- 18:02files that I've had in 2025. Imagine,
- 18:04for example, that you had to update
- 18:06this file every week. Well, for that,
- 18:07you have a PowerPoint here. This
- 18:09PowerPoint is right here, and it allows
- 18:11for analyzing all of Nike's e-commerce
- 18:14performance every week. You have to
- 18:16update it every time; it's boring, it's
- 18:18tedious, it's a lot of copying. Well,
- 18:21for that, a loop here would allow us to
- 18:23update it whenever you need it. For
- 18:25this, we go back to Cowork, select our
- 18:27folder, I choose a folder, and that's
- 18:29how I just added the Nike sales report
- 18:31here. And what I’m going to ask it to
- 18:33do is, every Monday at 10 a.m., update
- 18:36the PowerPoint presentation located in
- 18:38my folder using the data from the Excel
- 18:40files in the 2006 folder. So basically,
- 18:43it will open 2006 and take all the new
- 18:45data to update my slides. Imagine
- 18:48having to update your new Excel files
- 18:50every single time they change. You just
- 18:53have to drop them in, and then the loop
- 18:55will do the work for you. The best
- 18:57thing here is using a drive, because
- 18:59you don’t even need to upload files,
- 19:00and we get a truly 100%autonomous loop.
- 19:03Then, once the presentation is updated,
- 19:05it will generate the new version of the
- 19:07PowerPoint. It will go ahead and write
- 19:09an email for me. For example, we could
- 19:11have connectors or tools with Gmail
- 19:13here. In the body of the email, it
- 19:15presents the key results and prepares
- 19:17the email so it's ready to be sent. And
- 19:19from there, well look, every Monday at
- 19:2110 a.m., that’s what will allow for
- 19:23the full automation. And from there,
- 19:25we're off. And by having it set every
- 19:28Monday at 10 a.m., it sees that it's a
- 19:30recurring task. So in the end, when
- 19:33it’s finished, we’ll have a loop
- 19:35running every Monday at 10 a.m. without
- 19:37me. It knows where the Excel files are,
- 19:40it knows where the PowerPoint is, and
- 19:41it knows it needs to send an email. It
- 19:44will run continuously, allowing for the
- 19:46creation of a first very simple loop
- 19:49right here. After a few minutes, if you
- 19:51look closely, it clearly understood, as
- 19:53we saw, that it is a recurring task. It
- 19:55analyzed the folder here and asked me
- 19:58for validation. Do I need to have this
- 20:00done every week? I gave my approval. It
- 20:03did the entire analysis, it has the
- 20:05exact structure of all the charts in
- 20:06the PowerPoint, and it knows how to
- 20:08update them. The script is running
- 20:10without errors, it successfully
- 20:12performed a test itself to check the
- 20:14results, and it has scheduled the task
- 20:16to run every Monday. It has analyzed
- 20:18the structure and has all the
- 20:20components. The next execution is
- 20:22scheduled for Monday, July 20th, at 10
- 20:24a.m. If I go back here to scheduled
- 20:26tasks, you can see that I have my
- 20:28weekly PowerPoint and email report
- 20:30scheduled every time. So, it could
- 20:33update my tasks like this, and I would
- 20:35just have to go present them, for
- 20:36example, at a steering committee
- 20:37meeting. If I want to, I can still run
- 20:40the task here once to see the result.
- 20:42Here, we can see the task has just
- 20:44started, which allows me to run tests
- 20:46and ensure the quality of the results.
- 20:48And so, as we can see here, it has
- 20:50finished updating the entire file with
- 20:52the 2026 data. So, it’s basically
- 20:54like it updates every Monday; it comes
- 20:56in and retrieves the information. And
- 20:58if you look here at the top right in
- 21:00the output, I have a PowerPoint that
- 21:02has been updated. We were on 2025, now
- 21:04we are on 2026, and all the numbers
- 21:06here have been updated, which would
- 21:09allow me, once again, to automate all
- 21:11this reporting. And then, as I said, we
- 21:14can have it generate a draft email with
- 21:16this document as an attachment and thus
- 21:19send the email with all the items, all
- 21:21automated. I’m just taking a very
- 21:24short break in the video, less than a
- 21:26minute, but it’s important for those
- 21:28who want to go even further in
- 21:29mastering AI. Alongside me, you’ll
- 21:31find two resources in the description
- 21:32of this video that can help you. The
- 21:34first is my guide, "The 7 Days of AI."
- 21:37It’s a guide I designed to allow you
- 21:39to integrate all the basics of AI in
- 21:41just 7 days with me. Each day, you
- 21:43receive an exclusive resource that will
- 21:45help you progress, whether it's
- 21:47creating your AI agents, understanding
- 21:49the best prompt engineering techniques,
- 21:51or getting access to my community of
- 21:53over 8,000 members. In short, you
- 21:55receive exclusive content by email
- 21:56every single day. To get access to it,
- 21:58this guide is completely free. You just
- 22:01have to come here, enter your email,
- 22:02check this box, and press this button.
- 22:05And from there, you’re off on the 7
- 22:07days of AI, and you’ll receive the
- 22:08content for the first day directly. But
- 22:11on top of that, I know many of you want
- 22:13to automate as many tasks as possible
- 22:15using AI. And as you can see in this
- 22:17video, I am truly in a position to help
- 22:19you. Well, for that, you'll also find
- 22:21the presentation link for my AI Master
- 22:22masterclass in the description of this
- 22:24video. This masterclass is a benchmark
- 22:27in France. It's the only masterclass
- 22:29that trains you in all the essential
- 22:31skills for the AI era. Prompt
- 22:33engineering, vibe coding, context
- 22:35engineering—you will master all these
- 22:37terms and know how to use them. And to
- 22:39use them, you must master the best AIs
- 22:41on the market. Well, for that, the
- 22:43masterclass also trains you on all the
- 22:45top AIs available. ChatGPT, Gemini,
- 22:47Claude, Perplexity, Microsoft Copilot.
- 22:50By the end, you will master all these
- 22:52solutions. But above all, right now,
- 22:53there is a huge revolution. Agents:
- 22:56inside this masterclass, you also have
- 22:57an entire module to help you create
- 22:59your first AI agents alongside me. You
- 23:02can also find testimonials from people
- 23:03who have participated in this
- 23:05masterclass on the presentation page.
- 23:07More than 32,000 people have
- 23:08participated since its launch. You
- 23:10should know that this masterclass is
- 23:11open to everyone, regardless of your
- 23:13technical level. Skill building is
- 23:15progressive: we learn at your own pace.
- 23:17You can pause and come back a bit later
- 23:19, just like a Netflix series.
- 23:21Personally, however, I commit to one
- 23:22thing: AI evolves extremely fast. You
- 23:24have undoubtedly noticed it. Well, for
- 23:26that, I commit to updating the
- 23:28masterclass every month with new
- 23:30content. And obviously, from the moment
- 23:32you join, you get lifetime access. So
- 23:35you benefit from all these updates. So
- 23:37don't hesitate to check out those two
- 23:38links in the video description. But I
- 23:40know you definitely want to go much
- 23:42further into loops. Well, for that,
- 23:44rest assured, I have planned a much
- 23:46more advanced scenario. We will be able
- 23:48to create an entire team of AI agents
- 23:50that will work in perfect autonomy
- 23:52inside a loop. And for that, I chose a
- 23:55concrete client case. This client case
- 23:57is Trade Republic. And for that, we
- 24:00will imagine together that we work in
- 24:02the marketing team at Trade Republic
- 24:04and want to create a loop of agents
- 24:06that works automatically to create our
- 24:08Instagram posts, including the
- 24:10accompanying text, on the most trending
- 24:12topics. For that, what have I built
- 24:15here? I built this system that you see
- 24:17right here. Inside this system, what is
- 24:20it that will allow us to trigger the
- 24:22loop? So, having the AI here every
- 24:25Monday at 9 a.m., I want a market watch
- 24:27on the finance sector to identify the
- 24:30most interesting IPO of the week or a
- 24:32major market movement, a crisis, an
- 24:35opportunity, or a political shift. All
- 24:38of these aspects are important to me
- 24:40for creating an Instagram post. Because
- 24:42if you look here at Trade Republic's
- 24:45Instagram page, they share Donald
- 24:46Trump's investments, market updates, a
- 24:49case study on a specific company like
- 24:51Samsung, or the oil market here. They
- 24:54cover very specific topics. Here, the
- 24:57stock growth of Take-Two, the producer
- 25:00of GTA 6, or new IPOs, like SpaceX or
- 25:03Anthropic for instance. In short, all
- 25:07these elements are what we will try to
- 25:10create using this system to have
- 25:12LinkedIn posts that look exactly like
- 25:14the ones here on specific topics, 100%
- 25:17automated. From there, ultimately, the
- 25:20loop we're going to create will allow
- 25:22us to post on Instagram automatically,
- 25:24without me having to deal with it. It
- 25:27does this because I have an Instagram
- 25:29manager here; it's my planner, my
- 25:31orchestrator within the system. It
- 25:34distributes tasks to the sub-agents. To
- 25:37do this, it triggers every Monday at 9
- 25:39a.m. to launch the market watch. It
- 25:42calls the Finance Trend Agent each time
- 25:44, which is an agent connected to
- 25:46various tools. These tools include
- 25:49Perplexity. It will go to Perplexity to
- 25:52see which finance topic is most
- 25:54relevant in France to analyze and
- 25:56communicate about this week. So, for
- 25:59this, my orchestrator launches the
- 26:01agent. In the end, it has identified a
- 26:02topic. It then gives this topic to the
- 26:05copywriter. The copywriter is the
- 26:07expert in writing the Instagram caption
- 26:09with the right keywords, perfectly
- 26:11aligned with the chosen topic. And
- 26:14obviously, it replicates the exact tone
- 26:15and style of the posts we see on the
- 26:17Trade Republic Instagram. How did I do
- 26:20that? Well, you'll see, I'm going to
- 26:22train a skill called "Text DNA" that
- 26:24understands, based on all of Trade
- 26:25Republic's Instagram posts, how they
- 26:27are written. And finally, I have one
- 26:30last sub-agent specialized solely in
- 26:32image creation. For that, it has also
- 26:35been trained. It was trained on the
- 26:37Instagram posts that I scraped directly
- 26:39from the Trade Republic account. And to
- 26:42create the image, it has a tool, it has
- 26:44a nano banana that allows it to
- 26:45understand how it should create and
- 26:47have a good image generation tool. Once
- 26:50they have that, they are able to
- 26:52generate all the elements for me, they
- 26:53send them back to the Instagram manager
- 26:55who presents the result to me.
- 26:57Obviously, here, I chose to have the
- 26:59result presented to me. If we wanted to
- 27:02go even further, I could connect
- 27:03another tool that would allow me to
- 27:05post this directly to Instagram
- 27:07automatically, but here, I preferred to
- 27:09stop to be able to see the result with
- 27:11you before publishing it. And so, what
- 27:14we are going to do now is go and create
- 27:16this whole system together, and
- 27:18obviously, you can find all the prompts
- 27:20to create this yourself in my community
- 27:22using my free guide in the video
- 27:23description, the 7 days of AI. So, in
- 27:26order to create this whole loop system
- 27:28for Trade Republic here, automating the
- 27:30publication of my posts on Instagram, I
- 27:32already had to do several things. Right
- 27:34there, if we look closely, I created a
- 27:36folder called Trade Republic Marketing.
- 27:38And inside, if we look closely here, I
- 27:41retrieved some Instagram posts from
- 27:43Trade Republic. I took screenshots of
- 27:45them and we can see the post
- 27:46description on the right here and the
- 27:48type of image we want to reproduce on
- 27:49the left. And so, we are just going to
- 27:51need that to initiate the loop. And so
- 27:54now to initiate all of this, we are in
- 27:55Cloud Code. In Cloud Code here, you are
- 27:57going to select a project. For me, this
- 27:59project is very simple here, it's Trade
- 28:01Republic marketing. The first thing we
- 28:03are going to do is start by initiating
- 28:05the entire context of this project. For
- 28:06that, right here, I am going to send a
- 28:08first prompt. This prompt will allow me
- 28:10to provide all the instructions and the
- 28:12general context. I am the Trade
- 28:13Republic marketing team. The goal here
- 28:16of the agent system is to produce
- 28:18high-performing Instagram posts for
- 28:19Trade Republic. The structure, so I
- 28:22explain all the agents that I am going
- 28:24to put inside my structure. So, it's
- 28:26the little diagram that I presented to
- 28:28you previously. I explain to it that
- 28:29there will be a manager agent and then
- 28:31sub-agents. I also give it reference
- 28:34sources, the website, the Trade
- 28:36Republic Instagram account, the
- 28:37editorial objectives, the fundamental
- 28:39rules, which will allow me to provide
- 28:41the entire execution framework for my
- 28:43agent system. And here, as you can see,
- 28:46it has created this little .md file. So
- 28:48this allows you to start using it in
- 28:50Claude Code right here. But personally,
- 28:52I prefer to use it directly in Cursor.
- 28:55For that, you just come right here and
- 28:57select open in Cursor. From there,
- 28:59Cursor will open and you'll see we can
- 29:00use Claude Code without being a
- 29:02developer to follow through the
- 29:03scenario. If you look here, we already
- 29:06find what it just generated for me in
- 29:08Claude Code, and we can see the
- 29:10Instagram posts here, for example, with
- 29:12the different images since we selected
- 29:14the Trade Republic marketing folder.
- 29:17Now, we can ask Claude Code on the
- 29:20right to use all the Instagram posts in
- 29:23the "Instagram posts" folder, and also
- 29:26consult the .mmd file we just created
- 29:29with all the context, to generate a
- 29:32skill capable of understanding the tone
- 29:35, format, and atmosphere of the images
- 29:38—and only the images. Why? Because it
- 29:43will allow me to design a skill that
- 29:45can then tell subagents how images
- 29:47should be created. So, you launch
- 29:50Claude Code and it will start designing
- 29:52the skill for you. As you can see now,
- 29:54it has finished. The skill has been
- 29:56generated here. It’s called Instagram
- 29:59Visual DNA. Now, what I suggest we do
- 30:02is the same thing, but focusing only on
- 30:04the description text that accompanies
- 30:05each of the images in my Instagram
- 30:07posts. We send the request to generate
- 30:10the skill. Here again, the skill
- 30:12creation is finished. We now have
- 30:14Instagram Text DNA, which is a skill
- 30:17that allows me to generate this text.
- 30:20Now that we have that, if we look here
- 30:22and type/mcp server, we can see that
- 30:25several MCP servers are already
- 30:27connected. Indeed, I have already
- 30:30connected the MCP server with
- 30:31Perplexity, because if we go back to
- 30:33the copywriter system, we just created
- 30:35this skill. We also created the skill
- 30:38here for the visuals. Now, we will
- 30:41start creating the agent here for
- 30:43researching interesting finance topics.
- 30:46And for that, it needs to rely on
- 30:47Perplexity here. So, type/agent here,
- 30:50which will open the terminal, and from
- 30:52there, you will be able to create a new
- 30:54agent for a project you’re generating
- 30:56with Claude. You paste the information
- 30:59into it. Here, we want to create an
- 31:01agent specialized in researching
- 31:03current finance topics. What kind of
- 31:05finance topics are we talking about
- 31:07here? IPOs, growing or shrinking
- 31:10industries, key sectors, major
- 31:12political shifts; these are all
- 31:13important topics for us. Then, to
- 31:18conduct research, you must use the
- 31:20Perplexity MCP server, which will
- 31:22automatically search for information
- 31:25from the last three days to ensure the
- 31:27selected topics are current and
- 31:30relevant. We send the info, and the
- 31:32agent is being generated. Here, the
- 31:34agent is asking me which tools it can
- 31:36access. We’re going to give it all of
- 31:37them. It will run with a name, and
- 31:39we'll give it a little blue color to
- 31:41remind us of Perplexity. We link it to
- 31:44the project for memory. And here we go.
- 31:46The agent is created. As you can see
- 31:47here, Finance News Scout is the agent
- 31:49that was created. Now, let's ask it to
- 31:51create a second agent. This time, the
- 31:54goal is to tell it to create an agent
- 31:56specialized in creating Instagram post
- 31:58images. Its mission is to produce the
- 32:01visual, and only the visual. To do this
- 32:04, it will take the current news topic
- 32:06from our Finance agent, which will have
- 32:08analyzed the info in Perplexity. You
- 32:10never identify the topics to cover
- 32:12yourself. Next, I gave this sub-agent
- 32:14several steps to follow, but most
- 32:16importantly, it must respect the visual
- 32:18identity. For that, it uses the
- 32:21Instagram visual DNA skill we created
- 32:23previously. And to create the image, it
- 32:26automatically relies on the Nano Banana
- 32:28MCP. to create the image. I gave it a
- 32:31few constraints. It must only create an
- 32:33image and should never write text,
- 32:35draft a post, suggest hashtags, and so
- 32:38on. Its goal is strictly to make the
- 32:40image. I chose this deliberately. This
- 32:42way, I ensure that the final result is
- 32:44of very high quality. Now, as we can
- 32:46see, it has successfully created the
- 32:47agent. We can see the Finance News
- 32:50Scout agent here, and the Instagram
- 32:53Visual Creator here. So, I have my two
- 32:56sub-agents ready. I will need to create
- 32:58one last one. This one follows more or
- 33:00less the same process as the one we
- 33:02just created for the images. However,
- 33:04it will be focused solely on the text.
- 33:06This way, I have my sub-agent to find
- 33:08the relevant topic. And I have my two
- 33:10agents to create the image as well as
- 33:12the text. And once we've done that,
- 33:14only one step will remain, which is to
- 33:17create the planner, the orchestrator
- 33:19for these sub-agents, right after. And
- 33:21so now, I have my three sub-agents
- 33:23ready to go. If we look here for a
- 33:25second, we have this sub-agent that's
- 33:27ready. We have the other two here as
- 33:28well. They have access to their tools
- 33:31and various skills to ensure the
- 33:32consistency of the creations. Now, we
- 33:35are going to create the Instagram
- 33:36manager that will oversee all of this.
- 33:38For that, I come to the right, and here
- 33:40, if you look closely, you are the
- 33:41manager of my project's agents. We
- 33:43explain his entire mission to him. We
- 33:45tell him that, above all else, he must
- 33:46systematically check the folder. And
- 33:48here, we'll put trade republic
- 33:50Instagram agent.md with all the context
- 33:52. The goal is to create Instagram posts
- 33:54for Trade Republic. I've put the entire
- 33:57process here for him. First, he
- 33:58launches the Finance News Scout agent
- 34:01to go find relevant topics. Next, he
- 34:04creates the visual, and then he creates
- 34:05the text. At the end, he provides me
- 34:08with a full summary, the selected
- 34:09topics, visual concepts, the caption,
- 34:11and an explanation. I'm launching the
- 34:13creation of this last agent. Now, it's
- 34:15finished. We can see that it has
- 34:16created all the different agents for me
- 34:17. Now, what interests us is seeing if
- 34:20they are capable of working together
- 34:21and providing me with a good result.
- 34:24For that, I am starting a new session
- 34:25here. I ask the manager agent right
- 34:28here, the Instagram manager, and I tell
- 34:30him to start the process for the next
- 34:33Instagram post. And here we go. You can
- 34:35see that the full cycle has started.
- 34:38The first thing it just did, as I
- 34:39requested, is read all the context in
- 34:41the md file. And then here, it's
- 34:44calling the Scout agent to go perform
- 34:46the analysis on Perplexity and pull out
- 34:49an interesting topic for an Instagram
- 34:51post. You can see, by the way, right
- 34:54here, I can see the whole cycle, all
- 34:56the steps that will need to be
- 34:57triggered by the manager agent until
- 34:59reaching step 4b to output the entire
- 35:01Instagram post. And now here, after a
- 35:04few minutes, we see that the agent has
- 35:06completed all the tasks. If you look
- 35:08here, it has already tracked its entire
- 35:10to-do list. It went to retrieve the
- 35:12information here using the finance
- 35:15agent. It has brought back a topic.
- 35:18Then, it proceeded to use copywriting
- 35:20to write the post and the creative
- 35:21agent to generate the image. The topic
- 35:24it chose was the IPO of the brand Le
- 35:27Slip Français. So, this is going to be
- 35:29extremely interesting. Plus, it's
- 35:30French, so that’s really cool. And
- 35:32we're going to take a look at the
- 35:33result it generated here. In the end,
- 35:35if you look closely, it made all the
- 35:37edits and approved an Instagram post
- 35:39for the next cycle. Here, the Le Slip
- 35:42Français IPO is taking place on July
- 35:4414th; at the time I'm filming this
- 35:46video, it's the 17th, so we’re well
- 35:48within the three-day window that was
- 35:51set to create a small carousel. It
- 35:53prepared the first image for me, and
- 35:55then it prepared five prompts for the
- 35:57next visuals if I need them. It
- 35:59provided the caption with the text to
- 36:01be filled in. It also included a
- 36:03virality potential score. Now, we’re
- 36:05going to see what this image looks like
- 36:07, because we already have the text
- 36:09right here. Here is the image. It
- 36:11perfectly respects what Le Slip
- 36:13Français does. We have the blue, white
- 36:15, and red IPO flag here; Le Slip
- 36:17Français has just gone public. Here is
- 36:19how. And here, it really captures and
- 36:22reflects the DNA and the structure of
- 36:23the posts we have on Trade Republic. If
- 36:26you look closely, we have the large
- 36:27text at the top with a bit of text
- 36:29right below it every time. Here, it
- 36:31provided it in French for us. There you
- 36:33go, we have the same patterns every
- 36:34time. Well, not on these posts here,
- 36:36but here for example, we have all the
- 36:38right elements every time. The text is
- 36:41ready to go as well. It would just need
- 36:43to be published. Now, the last thing
- 36:46left to do is to reopen Cloud Code here
- 36:48and tell it: "Make sure this folder and
- 36:50the agents running in it act as a loop
- 36:53to deliver the post every Monday at 9
- 36:55a.m." And from there, now that I've
- 36:57tested the loop and it works, obviously
- 36:59, you could still make changes if
- 37:01needed. That's this testing phase. But
- 37:04here, let's imagine I want to continue;
- 37:06well, I can set up the loop. And from
- 37:09that point on, I don't have to do any
- 37:10more work at all. This agent system
- 37:13will run directly in a 100%autonomous
- 37:15way. And there it is, right here. After
- 37:18a few minutes, it’s all set up. Here
- 37:19is what was created. He created all the
- 37:21automation. Here, if you look closely
- 37:24on the left, I now have automation with
- 37:26the weekly prompt here, which lets me
- 37:28start the loop without me. I don't have
- 37:31to type prompts anymore. And here, we
- 37:33have the entire loop script that is now
- 37:35able to run without me. And so here,
- 37:39well, we managed to create a 100%
- 37:40automated loop together in just a few
- 37:42minutes, which would allow me to
- 37:44generate Instagram posts. Once again,
- 37:47if you want to go a little bit further
- 37:49here, you add a tool capable of posting
- 37:51directly to Instagram for you or doing
- 37:54it via computer use, as is already
- 37:56possible with Claude, and from there
- 37:58you are on a process. using the loops.
- 38:02So, normally with this video, you have
- 38:04everything you need to create your
- 38:06first loops and automate even more
- 38:07tasks with AI. And as you saw in the
- 38:10video, inside the loop, we created a
- 38:11small team of AI agents. But perhaps
- 38:14you want to create a much more complete
- 38:16team of agents using Claude. Well,
- 38:18that's perfect because right here you
- 38:20have a video displayed where I show you
- 38:22how to create your first team of agents
- 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.