In this guide
  1. Before you start: what you need
  2. Step 1: Build the scenario and add a trigger
  3. Step 2: Choose a connection and model
  4. Step 3: Write instructions and map the input
  5. Step 4: Add tools (and keep the list short)
  6. Step 5: Set the limits before you go live
  7. Step 6: Add a fallback connection
  8. Step 7: Pick a response format
  9. Step 8: Add an error handler
  10. Step 9: Test in chat, then run once
  11. Human approval: what Make offers today
  12. How an agent uses credits
  13. When credits run out (and the “paused with credits left” question)
  14. When responses get cut off (token truncation)
  15. FAQ

TL;DR: In Make, an agent is a module (Make AI Agent (New) > Run an agent) inside a scenario. You give it a trigger, an AI provider, instructions and tools, then test it in the built-in chat. Before going live, set a limit on Steps per agent call, pick a fallback connection, choose a response format, and add an error handler. Credits are charged per operation, plus AI tokens if you use Make’s AI Provider.

Disclosure: The Make link on this page is an affiliate link: we may earn a commission if you sign up through it, at no extra cost to you. It doesn’t change our advice, and Make has a free plan. See our affiliate disclosure. Method: this is a sourced walkthrough, written only from Make’s own help pages, app reference, pricing page and release notes, read on October 18, 2026. We haven’t built this agent ourselves, so nothing here is a test result. Make says the app is in open beta and that “product functionality and pricing may change.”

Before you start: what you need

  • A Make account. The Make AI Agent (New) app “is available on all plans using Make’s AI Provider.” Using your own OpenAI, Anthropic Claude or other provider connection needs a paid plan. (Make AI Agent (New) app)
  • A budget in credits. The Free plan includes 1,000 credits a month. Core starts at $12 a month billed annually for 10,000 credits a month (Make’s pricing page, checked October 18, 2026). See Make’s plans (affiliate link)
  • A small, clear job. Make’s own planning step asks you to write down what the agent does, its tools, its knowledge and its trigger before you build. (Create your first AI agent)

A good first job for a small business: read new inquiry emails, sort them, and draft a reply for a person to send. Keep anything that sends, pays or deletes out of the first version (see the approval section below).

Step 1: Build the scenario and add a trigger

In Make, “your agent belongs to a scenario.” Click Create scenario, add a trigger module (for example a Gmail or Google Sheets module), and set how often it runs. Only instant triggers can start the scenario as soon as new data arrives. Then click the plus icon after your last module, search for Make AI Agent (New) and add Run an agent. (Create your first AI agent)

Step 2: Choose a connection and model

In Connection, pick an AI provider. On the Free plan, choose Make’s AI Provider. On a paid plan, you can also use your own provider connection with an API key. Then choose a Model. Make notes that models differ in “processing speed, reasoning abilities, token cost,” and that Anthropic adds a system prompt for tool use, which increases credit use. (Create your first AI agent)

Two settings control how many tokens each answer can use:

  • Reasoning effort: “Higher reasoning improves quality, but uses more AI tokens.”
  • Maximum output length: a percentage of the model’s maximum. “Lower values use fewer AI tokens, but may cut responses.”

Step 3: Write instructions and map the input

  • Instructions: “Describe what the agent does, including its role and workflow steps.” The agent follows them on every task.
  • Input: the task or data for this run, mapped from earlier modules (for example the email subject and body).
  • Input files (optional): only with Make’s AI Provider, OpenAI, Anthropic Claude or Gemini, and a model that accepts files. Make lists JPG, PNG, GIF and PDF as input types.
  • Conversation ID (optional): set one if the agent should remember a thread (for example, map the email thread ID). “If you leave this field blank, your agent generates a unique ID for each scenario run and has no memory of previous communication.”

Source: Make AI Agent (New) app.

Step 4: Add tools (and keep the list short)

Hover over the plus icon of the Run an agent module and click Add tool. Make lets you add four kinds of tools (Create your first AI agent):

  • Modules for one-step jobs, such as creating a draft or updating a sheet row.
  • Scenarios (Call a scenario) for multi-step jobs. If the scenario needs to send data back, it “must end with a Return outputs module,” and you write a Description so the agent knows when to use it.
  • MCP servers (Add MCP) for tools Make doesn’t have. Make warns that “multiple MCP servers and tools may result in high AI token usage.”
  • Other agents as sub-agents. Nesting is limited to one level.

On the route between the agent and a tool, Tool settings > Tool outputs lets you choose which fields go back to the agent. Returning fewer fields means fewer tokens. To check what a tool returned before the agent acts on it, see our verification prompt for agent tool outputs, which includes Make setup steps.

Step 5: Set the limits before you go live

These settings are in the Run an agent module, under advanced settings. All quotes are from the Make AI Agent (New) app reference.

  • Steps per agent call: “the maximum number of times the agent calls the model per request.” This is your main guard against an agent that loops. Make doesn’t publish a recommended number, so start low and raise it only if runs fail for lack of steps.
  • Step timeout: how long the agent runs in each step “before it fails.” The default is 300 seconds and the maximum is 600.
  • Maximum conversation history: how many replies the agent remembers in a conversation.
  • Enable compaction: compacts the conversation “when it exceeds the context window.”
  • Enable fallback connection: see the next step.

Step 6: Add a fallback connection

Make added fallback connections on October 9, 2026. If the primary AI provider or model fails or is unavailable, “Make retries the run once using the fallback connection.” Make’s AI Provider is the default fallback, and the feature is on all plans. The same release added prompt caching: you pick a caching duration for supported Anthropic Claude models, and supported OpenAI models reuse prompt content automatically. Both reduce credit use, according to Make. (Make release note, Oct 9, 2026)

After a run, the module’s Reasoning tab shows “whether a primary AI provider connection failed and a fallback connection was used.”

Step 7: Pick a response format

  • Text returns a plain answer you map into a later module.
  • Data structure returns “a custom format, either as output items (Add item) or as a content type, such as JSON (Generate).”

For a triage agent, a data structure (for example category, urgency, draft reply) is easier to route with filters than free text. (Make AI Agent (New) app)

Step 8: Add an error handler

A step can time out, a tool can fail, and a model can hit its limits. Make’s help center lists five error handlers: Skip, Retry, Resume, Commit and Rollback. Retry stores incomplete runs so they can be retried automatically or by hand, and Resume sets a substitute value so the scenario keeps going. (Make error handlers) Make’s pricing FAQ, which still uses the older names Break (now Retry) and Ignore (now Skip), says error handler modules don’t use credits.

Step 9: Test in chat, then run once

Right-click the agent and choose Chat with Agent to send sample requests, check which tools it called, and adjust. Chat runs use credits too, including the operations of any tools the agent calls. You can also click Run once with data from an earlier run, disable one tool at a time to see how the agent behaves without it, and open History > Details to read the Execution steps and the Reasoning tab. (Create your first AI agent; Credit usage for AI agents)

Human approval: what Make offers today

The Make AI Agent (New) app reference lists no setting to require approval before the agent uses a tool. Make does have a separate Human in the Loop app that adds review requests to a scenario, but it “is available on the Make Enterprise plan” and “is currently in closed beta — only available to invited customers.” (Human in the Loop app)

On other plans, the practical option is to keep risky actions out of the agent’s tools. Have it write drafts or rows, and let a person send or approve them.

How an agent uses credits

From Credit usage for AI agents and Credits:

  • With Make’s AI Provider: Run an agent uses “1 credit per operation + credits based on AI tokens.” Tools work the same way.
  • With your own provider connection (paid plans): “1 credit per operation.” You pay the provider for tokens directly.
  • Knowledge files cost credits when uploaded (embedding tokens, plus file-description tokens on Make’s AI Provider). You are billed for embedding once, at upload.

Make’s AI Provider converts tokens to credits at a published rate per model. For its Large tier (GPT-5 mini), one credit buys 3,616 input tokens or 452 output tokens. As rough arithmetic, a run that reads 10,000 tokens and writes 1,000 would be about 2.8 + 2.2 ≈ 5 credits for tokens, plus the operation credits. Make doesn’t publish how it rounds, and rates “are subject to change,” so check the Token usage summary in the module output for your real numbers.

When credits run out (and the “paused with credits left” question)

Make’s pricing page says that if you run out of credits, “your scenarios won’t continue to run, until credits are added again.” Incoming webhooks are queued up to your webhook queue storage and processed when scenarios resume. Scenarios that poll for data “search for records since the last time they ran successfully.” (Make pricing)

To plan around that:

  • Queue size scales with your plan: 667 webhook queue items per 10,000 monthly credits, up to 10,000. When the queue is full, Make rejects the extra data. (Webhooks)
  • Extra credits cost 25% more than the credits in your plan. On Core billed annually, regular credits reset monthly and purchased extra credits expire at each monthly reset. (Extra credits)
  • Credits aren’t the only limit. Your usage allowance also includes data transfer (5 GB per 10,000 credits), and Make says “exhausting data transfer can trigger an auto-purchase even if there are still credits left.” If your scenarios stop while the dashboard still shows credits, check data transfer and your payment method (“if your payment fails, your scenarios pause”), then ask Make support. (Make pricing; Extra credits)

When responses get cut off (token truncation)

Make’s reference describes two separate limits:

  • Output too short: Maximum output length is a percentage of the model’s maximum, and lower values “may cut responses.” If a JSON data structure comes back incomplete, raise this first.
  • Context window exceeded: “You get an error message when the context window is exceeded.” Make’s fixes: upload a smaller file, upload the file as a knowledge file instead, reduce or change the tools, or lower Maximum conversation history. The Reasoning tab shows context usage, for example “575/400.0k tokens (0.1%).”

Source: Make AI Agent (New) app.

FAQ

Is Make’s AI agent free? The app works on the Free plan with Make’s AI Provider, and the Free plan includes 1,000 credits a month. Agent runs use those credits.

Can I use my own OpenAI or Claude key? Yes, on paid plans. Make then charges credits per operation, and your provider bills you for tokens.

How do I stop an agent from looping? Set Steps per agent call and a Step timeout, and add an error handler.

Can a Make agent ask me before it acts? Not in the AI Agent app itself, per its reference. The Human in the Loop app is Enterprise-only and in closed beta.

Is Make the right tool for my first agent? Compare billing and approval features in our n8n vs Make vs Zapier guide, or see Make alternatives and Zapier alternatives.

Sources & method

Method. Last checked October 18, 2026. Sourced walkthrough from Make’s own pages only. We haven’t built or run this agent, so no testing is claimed. Affiliate links: Make only (marked). How we work.