🎥 Recording & Recap: Build your first Make AI Agent!

Michaela · October 10, 2025, 9:38am

Hey Makers

A huge thank-you to everyone who joined our Community Live Session!

It was amazing to see so many of you there. For those who couldn’t make it, or if you’d like to revisit what we built, we’ve got everything you need right here.

Full recording

Catch the full session with our expert, @Benjamin_from_Make , right here:

Session summary

In this session, Ben showed us how to build our first Make AI Agent from scratch. He created a “Police Detective” agent that automatically analyzes a list of alibis to identify a suspect in a fictional crime.

Here’s a step-by-step look at what the final automation does:

Reads a list of alibis from a Google Sheet.
Feeds the alibis into the Make AI Agent with a specific goal: identify the most likely suspect.
The AI Agent then independently decides which “tool” to use based on its analysis of the alibis.
If the suspect is unclear, it automatically calls a scenario tool to write AI-generated follow-up questions back into the Google Sheet.
If a single suspect is identified, it automatically calls a module tool to generate and send a detailed report via email.

Q&A from the session

Q1: Can an AI Agent call another AI Agent?

A: Yes! You can create a new tool or scenario that calls another AI Agent and have your main Agent call that one. In other words, you can connect and mix any number of Agents. If you need something to be done in a nondeterministic way, you can handle it directly through scenarios and the kind of setup I showed in the demo.

Q2: Is it possible for an AI Agent to keep a session between multiple calls?

A: Yes, it is. When using Run an Agent, there’s a Thread ID field. It’s a unique identifier that allows the Agent to keep track of previous interactions, similar to a session.

If you want to ask a follow-up question and maintain context from earlier calls, include the same Thread ID. The Agent will then retrieve the conversation history and continue seamlessly.

This works the same way as in the chat interface, where the Agent understands that your follow-up question relates to the previous one. I encourage you to test it out and have fun experimenting.

Q3: Is there a way to limit the number of Tokens the AI Agent uses?

A: Yes. One way is to use scenarios for deterministic steps rather than having the Agent handle everything through MCP calls. Each call that sends or receives data consumes tokens, so letting the scenario handle predefined tasks helps reduce token usage.

You can also optimize the tools you provide. For instance, when searching a database, don’t return all the data. Let the Agent specify what it needs and return only that subset.

We’re also preparing an Academy course that will cover more tips and tricks for optimizing token usage in Agents.

Q4: Are the Make AI Agents able to implement deterministic flows?

A: In a way, the answer is yes and no. Agents can handle some deterministic steps, like in the MCP example I showed, where it called specific services such as “get projects” and “create tasks.”

However, there are limits. The best approach is to decide whether each action is deterministic (it must always happen in a defined way) or nondeterministic (depends on context or decision-making).

For deterministic tasks, like creating a task in a specific workspace or project, it’s better to use a scenario. For nondeterministic parts, like deciding when to create that task, let the Agent handle them.

Think of it this way:

Scenarios = deterministic automation
AI Agents = nondeterministic, goal-oriented decision-making

They complement each other perfectly.

Q5: How can Make AI Agents help with managing and automating different parts of a business — from admin tasks to sales monitoring?

A: You can use Make AI Agents to automate and simplify a wide range of business operations, whether you’re a solopreneur handling everything yourself or managing a sales team.

For example:

Administrative tasks: Agents can organize your calendar, respond to inquiries, update spreadsheets, or generate reports automatically.
Sales and performance monitoring: Agents can track metrics, summarize results, and highlight key trends or follow-ups, so you stay on top of your team’s activities without manual checks.

For inspiration and real-world ideas, check out these blog posts:

Q6: How can I design my AI Agent in Make so that it doesn’t invent information? And how can I build real trust in its responses?

A: Great question. Spending time on fine tuning your AI Agent prompts, bringing clarity to tool descriptions, and identifying the right foundational models to leverage are important parts of ensuring that any AI agent doesn’t invent information. You will likely also want to consider using a mix of more deterministic steps – and not relying solely on AI agents to make all of their own determinations. Make is ideally suited to allow you to use a mix of deterministic and non-deterministic steps in a workflow.

Building trust in responses can be completed by setting up “evaluation” workflows that are on top of your production AI Agent workflows. The evaluation workflow will feed your AI agent a large set of AI inputs and evaluate how well the Agent did against providing the right output (typically based upon AI or human scoring benchmarks).

Q7: Can you create an AI agent that would enable you to verify if what you want to build is doable on your website? And even give you the basics?

A: By “website”, we assume you mean the Make product offering. In truth, this question can be answered by any existing AI chat interfaces (e.g., you can ask ChatGPT about Make and it will walk you through the basics). We are continually working to bring these experiences directly into Make as well.

Q8: What frameworks do you use to determine the best use cases for agentic automation vs automation with AI?

A: We actually appreciated the practical advice written by OpenAI on this topic, available here: https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf.

This guide includes advice to when to use an agent vs. other workflow approaches.

Q9: Would the new feature be an AI that helps create integration solutions with GPTMAKER, for example?

A: While we cannot speak for GPTMAKER, it is certainly possible to use AI to create an integration on Make to any application that has an API.

Q10: Can a Gemini NotebookLM be a Context file?

A: This is not currently possible with Make AI Agents – but that is a great feedback for us to consider.

Q11: Can AI Agent be connected to social media like LinkedIn, Facebook as tools?

A: Yes, it can be connected to any social media in two different ways:

You can build a scenario that connects to Facebook, LinkedIn, Instagram, or any social media, and detects new posts, new coments, new ad form submission, when any information is retrieved, the scenario can call the Agent and ask it to handle it
You can build a Tool (a scenario) that connects to the social media to perform any action: create a new post, answer a comment, get comments from any post, search specific posts, and so on. This tool can then be called by the AI Agent when it’s processing a request.

Q12: Tackling error handling when using AI agent modules - any tips to keep in mind?

A: During the Live Session, we didn’t have time to cover error handling, but this question is very important. Thanks for asking.

Errors can happen in multiple places: at the level of the Agent execution itself, in the scenarios that call the Agent, or in any of the tools the Agent calls. Let me break down each of the 3:

In the Agent execution itself, Make provides a module called “Run an Agent”. If something goes wrong, this module will fail. In that case, the processing stops and an error message is triggered. You can add an Error Handler and add compensating actions, based on the error type.
In the scenarios that call the Agent: Make AI Agents can be run any time, with specific data and instructions. For this, you need the scenario that uses “Run an Agent”. In case of errors, you can do the classic Error Handling so that you can mitigate them.
In any of the tools the Agent Calls: if the tool is a Make Scenario, you can mitigate any risk of error from stopping the execution of the tool leveraging the same Error Handling technique. But, most importantly, when using an Error Handler, you can customize the error message returned to the Agent. In that case, the non-deterministic nature of the Agent will allow it to decide what next steps to follow. For example, if the error is a rate limit (429) and you return that information to the Agent (eg, “an error occurred: too many calls, try again in 1 min”), the Agent can decide to call the tool again after 1 minute. It means that it’s a good practice to use an Error Handler in tools and return any relevant information that can help the Agent to decide what to do next.

Q13: How do you manage the results consistency in a non-deterministic flow?

A: The data in non-deterministic flows can still be structured. But, if it’s not, for example, if a tool returns a full text of information, or if, for example, the data to handle is an email content, it’s mainly the LLM used by the AI Agent that will interpret and pick all relevant information. When the Agent needs to call a tool, and this tool requires structured data, it’s again the LLM that will transform any non-structured data to the expected format.
When you create a new tool that expects a specific field format, the best practice is to explicitly describe it. For example, if a tool wants a “meetingDate” Text input, you can set its description to "Text containing the date of the meeting. Must be in this format: DD-MM-YYYY.

Q14:I have a use case where we want to compare the images on a webpage to the image in a folder. Could this be done with AI agent?

A: Yes, in that case, you need to provide a specific tool to the Agent that will get the 2 images (via URL or Data) and compare the two images, using AI or any other information. You could use the Make AI Content extractor to ask it to describe each image and use Make AI Toolkit to ask it to say if the 2 descriptions are similar. Or, if you get the relevant information, you could just compare the pixel sizes, the file name, the file size, or any other metadata. It really depends on the information you can get from the 2 images.

Q15: Will we be able to have an AI agent log in to a website (WordPress, BigCommerce, Shopify, etc..) So the agent can work in there? If yes, how will this work? Via API, a bot actually logging in (with our credentials), etc..

A: You can implement it already, but you need to build a Scenario tool that will call a scraping or robot App that will perform the login and browsing of the website. Make integrates with multiple scraping solutions.

Q16: How to improve the system prompt message and how to add specific connect tools + MCP?

A: Prompting happens in multiple places:

In the system prompt where you define the role of the Agent, the general rules it should follow, the tone it should use, etc
In the user prompt where you give the Agent a task/goal and specific rules for the goal
In the Tools description
In the input and output parameter descriptions

Most of them can be tuned to ensure the Agent understands what it’s supposed to do and how to do it. If you use Make scenarios, you can type your own descriptions, but if you directly call a MCP Tool, you won’t have this control. If ever an MCP Tool is not properly described, you can still call it from a Make Scenario (using the MCP Client / Call a Tool module), and then describe your scenario and its inputs/outputs. You can then do your own mappings in the Call a Tool module.

Q17: Why would the status module not go after the add to Google sheet module, and instead after a router? Can’t the other one fail but we would still get the All done response?

A: We put it in a different route because we are using an Iterator. Since the “Return Output” stops the execution of the scenario, putting it after the Google Sheet would make the scenario process only the first bundle (so just on Row added in the spreadsheet).
If there is any issue in the first route (for eg, an error in Add a row), the scenario will fail and return an error message, so the Return Output “” all done" will not be executed. The Agent will know something went wrong.

The goal of the Live Session was to show rapidly how Agents work, and this scenario we showed can be largely improved. For example, we could have used a Bulk Add Row, and thus call once the Google Sheet API (and in that case we could have put the Return Output in the same route). We could also have added an Error Handler in some critical modules to capture any errors and return a nicer error message to the Agent.

Q18: Can the AI agent be trained with documents such as laws?

In Make, you can import files to the AI Agent, which can be considered as the “latest learning data”. They are used to give context to the reasoning of the LLM. These documents are used during the text generation phase to customise the response and to give more specific context to it.

If you want to “train” the AI Agent with a lot of specific data, in order to give it a specialty, you need to use your own LLM, customised with your data, and you need to select this LLM when you configure the AI Provider connection (only possible from the PRO plan and above).

Q19: How to upload documents, and how many times does it need to be done?

A: If your question is about uploading a document so that the AI Agent can work on it and achieve its goal, then you can leverage the scenario that calls the Agent. You first download the document, and then you use “Run an Agent” and you add the document in the user prompt. You can also create a scenario tool that the Agent will call when it needs. This can be done any time, but since it consumes Tokens, you should monitor your consumption.

If your question is about uploading a Context File so that the Agent has internal information that helps it achieve its goal, you can upload them directly in the Agent configuration page, or you can use the provided modules in the AI Agent App, in a scenario. You may want to use the modules if your files need to be regularly updated. Be aware that it also consumes tokens (hence Make Credits).

Q20: For large Customer/Invoice datasets, what’s the best practice in Make to handle QuickBooks rate limits and pagination (e.g., iterators, backoff, cursors)? Can I do ‘modified since ’ incremental syncs natively to avoid full reloads?

A: Yes, you can do it directly in the scenario tools that you create. There are many different techniques that you can use to avoid any rate limits or to handle pagination.
The best practice is to use filters as much as possible. If the API you call allows you to use “” modified since’ you should definitely leverage it, and create a tool that expects this parameter. The AI Agent will then use the proper date when it calls the tool.

If you risk getting rate limits when inserting data, you can implement different techniques. For example, you can build a scenario tool that gets all data (from the Agent for example), and calls asynchronously a sub-scenario that will handle all the data with sleeps. In that case, the AI Agent will get an immediate response (so that it can continue its work), and the subscenario will handle the data on its own.

Q21: Is Make compatible with Gmail? I got a lot of error messages trying to get the flow successful.

A: Yes, Make is compatible with Gmail. We just released a new Make Gmail app version that helps connect to free Gmail accounts, you may want to test it. If you get errors, you can ask questions in the Make Community so that you can get help.

Q22: Can you give a short recommendation on how to use scenario input /output values?

A: Each scenario input and output should have a well-defined type (text, boolean, date, array, collection, etc), and be clearly described. The description is the most important asset that the AI Agent leverages to understand what it means and how to use it. For example, you can explain the format you expect, you can say when to use or not use the parameter, etc.

Q23: What is the best way to introduce my make AI agent into my WordPress website?

A: We would need a more precise question to answer properly, but globally, Make integrates well with WordPress. We provide an App with many modules that allow handling posts, media, users, categories, comments, etc. You can, for example, build a Make Scenario that gets the latest post comments and calls the AI Agent so that it handles them. Or you could use an AI Agent to make it generate a new Post, and then use the Make App to push the new post to your WordPress.

Q24: When to decide to use agents that can also act as tools?

A: You create and use an Agent when you need to implement fuzzy logic / non-deterministic flows. So, if a Tool requires such automation type, you will use an AI Agent in the tool.

To give you an example. Imagine you are creating an AI Agent that handles all incoming customer queries (from different places), when it’s a nice comment in a social media, you want the AI Agent to simply answer, when the comment is bad, you want the AI Agent to send an alert to the Customer Experience team, and when it’s someone that requests help or found a bug, you want the support to handle it. In that case, your AI Agent could call a tool, passing the request. And in this tool, you could have a second AI Agent, specialized in Support Requests that will define the criticity, the relevant support engineer, create the relevant ticket, and perform any follow-up action. In this case, the first AI Agent has delegated a specific task to a second Agent.

Q25: Now I’m using now OpenAI Chat Completion module to make AI processes in my scenarios. Is it a good idea to go for Make AI Agent instead of the OpenAI module?

A: If you just need to do a completion, you can keep using the OpenAI module. If, later, you need to perform different actions based on the data you get, you may want to use an AI Agent. But as long as it’s a simple process of unstructured data, a simple AI module is enough.

Q26:Can Make pull information and photos from my blog, i.e. a photo and its description? (I mean, without my having to set up a Google Sheet with the photo’s URL and description in it?

A: It will depend on what blog technology or what cloud drive application you are using. Make can call any API to grab information about files, download the files, analyze them (using AI), get their description, but all depends on the presence of the API and its capabilities. For example, it works very well with Google Drive, Dropbox, Egnyte, Box, but also with blogging technologies, such as WordPress, Wix, …

Q27: If I give access to Make to my Gmail and my Drive, how do I make sure that there are no data leaks from my email? How robust is Make’s cybersecurity?

A: Make is now SOC-2 compliant, and we regularly publish online documents about the security measures we take as well as how we handle data. See here.

Besides, in your scenarios, you have the option not to keep any data in the execution log. See here

Helpful resources

Help Center: Make AI Agents
Make Academy: Automation to AI Agents: Foundation
Blog: The difference between AI agents and agentic automation
Blog: What is agentic AI?
Blog: The automation journey: From traditional automation to AI agents
Guide: How to determine when to use traditional automation vs. AI agents
Guide: How to build AI agents with Make

We’d love to hear your thoughts

What was your biggest takeaway from the session? Share your thoughts or ask follow-up questions below.

Topic		Replies	Views
AI Agents in Make Features	4	1529	June 19, 2025
AI Agents + Grounding Features ai , ai-agent	2	123	July 30, 2025
Make AI Agent module Features ai	2	126	July 21, 2025
Alice - an app that runs Make scenarios with AI 👀 Showcase api , chatgpt , ai	6	913	August 26, 2024
I need a Make Pro who can help me build an AI Agent (+ educate me how it works) Hire a Pro professional-service	3	186	March 3, 2025