Agents Workspace ChatGPT : la vraie automatisation pour Tous enfin !

Do you know what the number one cause of layoffs is? Poor performance these days. Producing faster, better, and cheaper is essential for every company. And if you're an employee, the employees who will be retained are those with the skills to use and configure artificial intelligence. Today, it's absolutely crucial that you know how to use ChatGPT's Workspace Agents. It will allow you to automate part of your daily work with artificial intelligence. And that's precisely the goal of this tutorial. We'll start with something simple so that by the end of this tutorial, each of you, each company, will understand the principle of automation. And if you have a ChatGPT Plus account, you'll all be able to create your first automations by the end of this video. I'm going to show you how we're going to redo the OpenAI tutorial. We're going to improve it. Create a system that reads your appointments, updates them, generates client profiles, and reminds you what to do during client meetings. Imagine a system that can assist you every day so you can be more efficient, work better, and be more productive. That's exactly what we're going to do in this tutorial. ChatGPT Agent is the latest feature in ChatGPT 5.5, and it appears on the left. For a few days, it will be completely free until May 6th. The agent section, located right here, will allow you to create AI agents. So, what is an AI agent? It's a system that retrieves instructions, prompts, to perform actions—your work. To do this, it navigates within what's called context. For example, it might retrieve information about the client or your data. Therefore, you'll need to provide what are called connectors, and more specifically, MCP connectors. The model will therefore make decisions: which document should I retrieve, what should I do? And it will check the results before sending and validating them. So today we're going to use your first AI agent system for work. Here's how we'll proceed. In the sidebar, you have an agent system. This is what it looks like when you haven't created any agents. The steps to create an agent are as follows. You need to describe the automation you want to implement. And today, we're going to start with a really simple but effective tutorial, the one OpenAI did during the demonstration. What you need to keep in mind is that the process of generating an AI agent, the one that will work for you (the ChatGPT Workspace Agent), is a nine-step process. We have a section where we'll describe what we want the AI ​​to do, and we'll build the workflow with what's called the Workspace Agent's Agent Builder. You'll realize you're not alone in creating the system. They've designed an AI agent that will allow you to ask the right questions, connect the interfaces, and link the crucial actions—those of the MCPs (Methods for Communication Points). MCPs are tools that enable communication with other external tools. For example, if I want to use a Slack, Gmail, or Google Calendar account, ChatGPT Workspace Agents can read, write, and retrieve information. Step number 3 will therefore be adding the connectors. We'll add connection functions to allow ChatGPT Workflow Agent to retrieve and send information. Don't worry, it's simple to set up, and I'll show you how. We'll give these connectors permissions. Some will be able to read, others to write, and what we want is to keep the system secure. An AI agent is only effective if it writes as you intend. Since the company has its own graphic charter and its own way of interacting with customers, we're going to create a Behavioral memory. And that's called a "skill." Then there's the ability to maintain consistency between each action, and this time, that's "memory." The final step is that we'll test the agent before planning, executing, and deploying your agent—your first AI agent for ChatGPT Workflow Agent. How will we automate part of the work? These same skills that you sometimes pay a lot of money for on the AI ​​market, you'll be able to code yourself. If you want to go further in mastering AI, in 15 days, you'll be able to go from a simple ChatGPT user to an entrepreneur or employee capable of building useful, clean, and auditable AI systems that can be activated in the enterprise. You'll learn how to deploy, structure, and monetize your subscriptions. Imagine a ChatGPT subscription at $20 per month that will perhaps generate 30% more productivity for you. In this ecosystem, I 'll provide you with 30 pre-configured AI agents, tools, 80 hours of self-paced training, and all updates included. I 'll teach you the latest AI agent creation technologies to optimize your time and work. All the details are in the description. For this tutorial, I suggest we use OpenAI's Workspace Agents as an example to replicate the same system and then improve upon it. Here's what we'll build: a system that will retrieve our appointments, prepare them the day before by gathering all the necessary client data. We'll have a client profile, all the information needed to optimize the appointment, and we'll code this using ChatGPT Workspace Agent. Go to the "Agents" sidebar menu and click "Create an agent" in the top right corner. We'll use the prompt I'll provide in the course. The idea is to create an agent for sales meetings. We'll click the "Enter" button, and you'll see the different steps involved in creating the system. The interface is divided into two sections. The right side is the "Editor" system. If you click on it, you'll see that this is where the Workspace ChatGPT Agent workflow will be built. The system will be given a name, tools, the MCPs (Meeting Points) that will connect, skills, and information that can be downloaded from the "memory" functions. When we connect the system, from our prompt, the model will indicate that it needs to connect, and that's when we'll activate the MCPs. To retrieve our appointments, the model needs to retrieve information from Google Calendar. To retrieve the customer information we've stored, we'll give it access to our Google Drive database. To retrieve information about email exchanges with the client, it needs access to Gmail. So, we'll connect these three MCPs and begin building the system. We'll click "Create" and proceed to the next step. The Agent Builder shows you in real time what it's doing. I need to connect three MCPs. It integrates the three MCP functions, and in the next step, we'll activate them. I'll show you how in the "Sales Meeting Planner" section. That's the name given to the tool. You can change the name if you wish, and on the left, you can see the reasoning process of an AI model. What you see unfolding on the side is actually a model of the model's training data. It's not by chance that ChatGPT acts this way, that is, that it performs a list of actions; it's because it has received training. Human. Skills have been transferred to it, what's called RLHF, reinforcement through human learning. Thanks to this system, the model is able to anticipate and produce part of your work. But what you'll discover is that in automatic mode, the model remains an AI; that is, it will write like an AI, and you'll realize that the role, the workflow, or even the reasoning system wasn't designed by a person—it's still an AI. An AI is a succession of word probabilities. This means that often the AI ​​doesn't know what it's producing. It does it simply because it learned it from the training data. This is where, at some point, you'll intervene as an AI architect and optimize what the ChatGPT Workflow Agent is currently capable of. The Workspace is almost fully configured. To activate your connector system, click on the element and then simply click the "Connect" button. A window will appear, allowing you to connect directly to the system. This is a secure OAuth connection. Here's how: Click the "Next" button, connect your interface, and the system will automatically allow you to synchronize your account for both reading and writing. To understand an MCP system, in the "MCP Functions" section (Google Calendar GPT Workgroup Agent), scroll down to the entire section of functions. What you see here is the technical name of the MCP function. When you connect MCPs, all this information becomes part of the context. This means the agent knows it can use all these tools. You can configure the tools for an agent. Click on the right-hand side, "Content by the agent," and you will be able to modify access rights. Here, you can always specify "don't ask" if you trust the configuration you've created. The system will then be able to work completely autonomously. Alternatively, you can customize a security level. Let's say you don't want a system to be able to send an email on its own. In that case, you can configure it for open access or "ask before running." Of course, if you ask before running, the system won't be fully automated. You'll still have to press a button and be at the computer at some point. The very purpose of automation is to allow you to work even when you're not at your computer. ChatGPT will run an automation completely independently on OpenAI's servers. In our case, we'll enable everything and set "don't ask" everywhere. This allows us to avoid having to validate ChatGPT's actions during write operations. So, only do this when we reach step number 7. The tests show you that the model is functioning correctly. When you've configured a system and changed a model's behavior, you update it. The update will apply the new settings to the system prompt. When your agent is ready, you can click in the chat and switch to the testing phase. So, as you've seen, the setup was very quick. To return to the initial page, go back to the "Back" section. This is where you have the option to add skills. We'll need to think ahead a little to consider what will happen in the workflow. As we said earlier, our system will retrieve the appointments from the calendar and prepare a client sheet to remind us of the essential elements for the appointment. So, if we want the The system writes this form in a specific way, so we'll need to create what 's called a "skill." We'll add a skill, and we can either load this skill if you already have it, or create it. There are two ways to do this. Either we click on "Skill" and describe the skill. Of course, this is an introductory tutorial. So we're dealing with something simple that everyone will be able to do by the end of the video. We can describe a format, and the best way to describe the format is to give an example. If I want my meeting brief, the one I use for my company, to look like this—that is, a summary, meeting information, the meeting date link, client account information, and the client's positioning—well, the best thing to do is go back to the ChatGPT Workspace Agent interface and use either the "Create Skill" function to create a skill, or ask the Agent Builder, the Workspace Agent builder, to create the skill. The goal is to create a skill. The skill will be loaded when the meeting minutes are written. The meeting minutes will be applied to this template. So, we provide the information to the Builder and specify the function along with the template. We'll provide the data templates and let the Builder create the skill section that will add the skill. So, as soon as we start working on this type of system, you'll see that the entire building phase will require taking the time to understand what the system includes and what you get. Always compare it to see if it aligns with your objective. Therefore, always use the reasoning area of ​​the ChatGPT Builder agent, which will allow you to track whether the steps correspond to what you want in the end. The final step of the Builder system has the advantage of automatically linking the skill to you. Otherwise, you'll see that in the "skills" section, your skills will appear on the side panel, and you 'll be able to add the skill to your AI agent system. Incidentally, you might not know this, but there are already skills built into ChatGPT, and we can call them, which I'll show you in the second part of the video. We don't have to code everything. We can call upon certain skills that are correctly written and that we'll use later. We're going to anticipate a rather interesting point: memory. What is the purpose of memory? It's to maintain consistency over time. So, in the context of appointment or customer management, the idea is to tell ChatGPT how to use memory. While ChatGPT Builder creates the skill and uploads it—meaning it automatically adds the skill without us having to do anything else—it takes control of the screen. I don't do anything, and it sets up the skill system. I'll show you that we can then see how the skill is built and modify it if necessary. Now, just click the button and read the structure. This is the template: when to use the skill, its purpose, the workflow, the writing rules, the information management, and how the content is formatted. He then put the example in the "assets" section. So this is how we create skill systems. For those of you who have the courses, I'll provide the optimized template for creating skills. You'll have the template where you can simply copy and paste. This will allow you to create much more complete and optimized skills, and you can then integrate them directly into the interface. So what we've just done is added the skill, and now we're going to update our system. We'll click on the Clicking the "Update" button will validate the new skill in the workflow. The change is that we now have a skill, but we need to tell it when to use it. So, the system prompt update command is: brief writing step, skill function call. It will then search for the function, and we'll give it the skill name. We copy the skill name and assign it the name of the variable, and it will update the system. So, going back to this section, the key point is that we have a function, but it's not named within our system. Therefore, if you want the system to launch at the correct step, my first tip is to integrate the skill call into the workflow. The ChatGPT Builder system automatically identifies the section and integrates it into the workflow. Logically, it will go to the system prompt as instructed and update it using the "skill directory." As you write, you structure, and you update. So now, we've just updated the system and we're updating the block. The last step is to talk about memory. Regarding memory, what's important for maintaining consistency in the workflow is that the system knows if we've already addressed the topic and where the files are located. We're going to ask it to update the memory section. So, by clicking on the memory section, today the memory system is empty. What we're going to configure as the memory system is the ability to remember the client's name and indicate which directory contains the client's documents. "Activate function 'open quote memory closed quote'. When an appointment is confirmed, include in memory the client's name, the ASCII directory tree of client files from Google Calendar, Google Drive, and Gmail. Execute a memory update check step at the end of each command execution." So, at the end, it will check and update the memory system. We're going to update the memory section, and the advantage of using the Builder is that ChatGPT will code some of the instructions. But I want to come back to something quite important: the writing process. You should know that the way the AI ​​writes its instructions isn't optimal. It's fine for getting started. It will be very useful, but you'll see that if we use it in a more professional setting, we'll need to be more skilled at defining the model's parameters. The main problems we'll encounter are that we have far too many words and not enough definition of how the model works. We'll discuss this in future videos, but the goal is for you to have a reliable, auditable system that you can rely on while the AI ​​works for you. So, it will still integrate the first part of the system's operation, which means it will automatically update the system prompt. We have a memory function and memory sections with the workflow description. Again, we'll update this. Once updated, we can start testing. We've finally reached step number 7. For testing, I recommend trying it directly in ChatGPT by clicking the "Try in ChatGPT" button. If the tests are successful, as I'll show you right now, you'll be able to schedule when the system starts. Since we've asked it to start every day the day before, we'll have a system that runs automatically every day at 6 PM. But you can delete, modify, run a test, or even define when you want to run this workflow. You can validate it by hour, day, and define the operating time zone. We can also provide instructions. Additional information may be needed if there's specific details. If you have multiple collaborators, each might want to personalize certain elements, and you can integrate that into this section. Now we'll move on to the testing phase. To do this, go to the main interface. I'll show you where to find your agents. In the left-hand section, you have the "Meeting Preparer" section, and this is where you'll be able to test your system. You'll have access to its activity. You'll be able to track all the times it has worked, the applications it has, and the permissions it has. By clicking on it, you can access how the model works. Regarding the Google Drive view, you saw that we have the option to let it act without asking it to validate actions each time. So, for testing purposes, I'll show you how to put the system in automatic mode. It will make the decision and then execute the task. To do this, click on the right-hand button, "Edit Agent." This will be an opportunity to see how we modify the actions. We'll go to the Google Drive section. You'll click on the "Google Drive app MCP," the agent's account. In the "Write action security" section, you 'll switch to either "Customize" or "Never ask." Of course, you should only do this if you trust the configuration you 've made. We're updating our system, and the last MCP to set up is the Gmail section. And we're going to do exactly the same thing. This will allow, if you wish, the model to work independently on the writing aspect. When we talk about writing, we mean the ability to write information. This means that at some point, you'll be able to send emails and schedule appointments with your agent system. And that's where it starts to get interesting. So, we'll update our system and then we'll test it. Now that we've seen how to modify the operating parameters, let's try it out by clicking "Try in ChatGPT." What we'll ask it to do is check if we have any appointments for tomorrow morning. "Checks appointments for tomorrow morning." By default, the system will deduce the time and time zone because we've already provided them, and it will retrieve the information from the Google Calendar scheduling system. Now, one thing to keep in mind: you would be much faster than artificial intelligence. You would have immediately gone to check your Google Calendar. AI isn't designed to be ultra- fast; it's designed to perform a task by following specific steps. That's what we explained at the beginning. We created what we call a workflow. It will check the calendar events, read the instructions (how we read the files), and see if there are any appointments for tomorrow. From there, it will check if there's an appointment, "how I should behave." So, we've integrated the behavior into the system. If it detects an appointment, it will retrieve the information from Google Drive and check if we have data on the client. From there, it will prepare the summary for the client meeting. And for that, it will use the skill. This skill will format the document, and at the end of the section, it will remember that it updated the memory because it just found the client information, and next time it will be even faster because it will know where the files are. So, it found an appointment for tomorrow morning at 10:00 AM Paris time with the company Parlons IA. It took 1 minute and 19 seconds to perform this check. So, you understand that the system's principle was respected, and I 'll show you what it did while we looked at its workflow. The system then read the record that was already in the database and prepared the meeting brief for tomorrow. "Tomorrow, I have a meeting with the company Parlons IA, and thanks to ChatGPT Workgroup Agent, I have a summary of tomorrow's meeting. I have the meeting information, the meeting link, and I can click on it if needed to check the appointment scheduled for tomorrow morning. So I have a functional meeting summary system: the position of the offer, the meeting objectives, the points I need to pay attention to, the questions to cover, the missing information, and all of this is because the model was able to build it within the skill." If you've reached this stage, congratulations! You've created your first automation, retrieved information from your meetings, created records, and now we're going to create a system where we can send emails. In the interface we have now, we're going to make some changes. We'll go to the "Edit Agent" section and add some functions. The idea is to modify our workflow. The goal is to modify the workflow. Send an email with the subject: "Appointment Reminder," the appointment slug, the appointment date, the appointment time, and the message body: a copy of the brief. We're going to integrate new information with an action on the Gmail MCP function. We'll activate the "Send Email" MCP function. If you don't know where it is, you'll see that by clicking on the Gmail section, in the "Functions" section, you'll find a section called "Send Mail." This is the function we're going to activate. We'll then apply the modification we need to prepare. And now, the goal isn't simply for him to write it to me in Google Drive, but for me to receive the information and be able to immediately recognize the email because it's an appointment confirmation. I'll receive it every day at 6 PM and I'll be ready for tomorrow morning. So what he's going to do is take back control of our prompt system and add a step. In addition to creating a record with all the client's details for the appointment date, we'll receive it by email. Since the MCP is already connected—we've already done that, we've already enabled the sending option—the entire workflow will be completely automated. We'll simply allow the system to communicate with us. We won't need to go to Google Drive to retrieve the information. We'll simply check our personal email inbox. Here's how the model interpreted the information. He understood that my system will retrieve the appointment reminder with the appointment slug, date, and time. Then, we'll have a portion of the message body that's a copy of the brief. So we have two parts in the command section. The system is now taking control of the screen. ChatGPT is modifying the role's behavior. It's showing us that it will customize the AI ​​agent, ChatGPT Workgroup Agent. We'll just read through it to verify how it will work. It will now use Gmail to send an email at 6 PM, aligned with this new format. So, it will use Gmail to retrieve relevant exchanges for the client and send the appointment reminder emails. We have the functions integrated here. For email, we have a formatting section where, via the Gmail function, you have the defined structure and composition rule. And then the information is updated in the memory section. So, regarding the instructions, I assure you we can do much better. But to get started, I think this will already help many of you to create workflows to optimize operations. So we're going to launch a query and We'll see if we receive the email for tomorrow morning's appointment with the information. And if there are things that don't work, don't worry, that's what we call an iteration. That's when we intervene, and if it works, I'll show you how to share it with your teams right after. "Check tomorrow's appointments. If there's an appointment, send me the brief to my email address." So we send them the request. Remember one more thing, as I said, an agent system is n't meant to be fast, but to follow steps and therefore automate the same processes every day. Imagine you're a real estate agent; you get inquiries about properties or products every day, and you can automate the responses for each property, each house. You can create a dataset where the model will answer almost 90% of the questions, and you can create a system that will handle all the drafts. Here, I chose sending because it will show you that it's capable of sending. But overall, you could actually prepare all your emails with artificial intelligence and go to the final step: manually validating them. And you can absolutely do that. You just have to modify a few sentences, correct them if necessary, and send them to the client. And when you see that your system is stable enough, you can send them automatically. So here we have the system that retrieves information about Paris schedules. Checking the calendar, checking the Gmail profile. Are there other items in my inbox? It follows the workflow. Now, it reads the events. It will check if there is any data in Google Drive concerning the client. We might have stored old appointments, old reports. And it retrieved information about the client. It retrieves the data. So it retrieves the context. Remember that agentic systems, as we said, are always based on the same pattern. A prompt triggers tools, the tools retrieve context, and then it triggers behaviors, and it's up to you to learn how to code all of that. So it updates the memory, and we'll see if we received the email. So what does it tell us? "For Monday, you have an appointment at 10:00 AM Paris time with the company Parlons IA." So the ChatGPT Workflow Agent did retrieve the information. Now, what it told me is that, according to the operating conditions, the participant didn't confirm their attendance. So, according to the rules, I didn't send the brief. So, what we're going to do is ask them to send it by email anyway, but that's very interesting. This shows you how closely models are what we call "aligned" with instructions. So, we hadn't really paid attention to the fact that there was an instruction that said, "If the appointment isn't confirmed, I won't send the brief." Okay? So, we're still asking it to send it. It sends a "send reminder email" to the inbox, and we'll have all the information. We've set up a system that demonstrates one thing: models need clear instructions and the ability to handle conditions. What we call conditions are "if... then..." statements, and what happens if something happens. So, while the model finishes, sends the information, and updates the data memory, I'll talk to you about conditions and prompts right after this, and I'll show you how we share all these elements. So the system confirms that it sent a "Let's Talk AI" slug at such and such a time, client appointment with the appointment time, and it updated the memory with the information, and I confirm the appointment by email, and I have all the information that is copied from the data we just indicated. For now, everything is working perfectly on the system level. If you want to share with your teams, you can click the "Share" button, and the sharing links will only be visible to members of your workspace. They can also customize them. If you go to the "Edit Agent" section, you can create what are called public agents. So, "I accept connection sharing." Anyone with the link will be able to chat with your agent or remix it to create their own. This means you'll be able to share instructions, not connections. So, if you share the system, remove your login credentials from the interface. This is very important. Remove your MCPs if you're sharing it externally. You can do this by clicking here, then click "Disconnect." Don't share publicly, only with your company's team, the workgroup that belongs to the same ChatGPT workgroup. So, in the memory file, you have the option to click the "Memory" button and check the information that has been collected. When you click, you'll likely be able to access and download the file. And you have exactly what we asked the AI ​​for: a data tree structure showing where the files are located in memory. This will allow the model to retrieve customer information much faster next time. If you've made it this far, congratulations! You've created your first automation system from scratch. You've learned how to modify your AI agent's functions. Take advantage of it; connections are free until May 6th inclusive. Now is the perfect time to expand your agent fleet and automate your work. In the next videos, we'll set up a more technical system where we'll learn how to build workflows like true professionals.