How to build an AI agent for free: method, tools and examples

The word "agent" now gets stuck on any chatbot. To build one that is actually useful, you first need to know what it is, then follow a simple method. I start with a short definition and seven examples, then walk through how to build an AI agent step by step with free tools, and I finish with the limits you need to know before you hand it anything.

What is an AI agent?

An AI agent is a program that pursues a goal on its own: it decides the steps, uses tools (files, web search, email, spreadsheets), checks its result and tries again if needed. What sets it apart from a regular assistant: you give it a goal to reach, not an instruction to execute word for word.

AI agent vs chatbot vs automation: the difference

A chatbot stays reactive: without a message from you, nothing happens. A classic automation is perfectly predictable, which is its strength, but it stops at the first case nobody planned for. An agent adapts to what it finds. It is more flexible and less predictable, and that is exactly why it needs guardrails.

The four components of an AI agent

• A language model (Claude, GPT, Gemini, Mistral), which reasons, decides and writes.
• Tools, which let it act: read a file, search the web, query a database, write to a spreadsheet, draft an email.
• Memory: the context of the current task, and sometimes preferences kept from one run to the next.
• A control loop: perceive the current state, decide the next action, act, check whether the goal is reached. If not, start again.

Remove the tools and all that is left is an assistant that talks. Remove the control loop and you get an executor that does not know it got things wrong. To connect an agent to your software, the standard today is the MCP protocol, published by Anthropic in late 2024: I explain it in what is an MCP server.

7 real AI agent examples

Here are the seven agents I build most often. For each one: what it does, and the simplest tool to build a first version.

1. The competitive intelligence agent

Every morning it goes through a list of sources, spots what is new, filters out the noise and sends a ten-line summary. It is the easiest agent to build, because you can judge the result immediately. To build it: n8n or Make with a scheduled trigger, an RSS feed reader and an AI step. I walk through it step by step further down.

2. The AI citation tracking agent

Every week it asks answer engines a panel of questions, records who gets cited and with which sources, then feeds a dashboard. It automates the method described in my article on the AI visibility audit. To build it: a small program or an n8n workflow that calls the model APIs.

3. The inbound request agent

It reads incoming form submissions, identifies the type of request, fills in the record in the CRM and drafts a reply that the salesperson approves before it goes out. To build it: Zapier Agents or Make, connected to your form and your CRM.

4. The pre-publication check agent

It checks every page before it goes live: tags, broken links, oversized images, structured data. It returns a table of issues. To build it: a reusable skill in Claude (see my Claude Skills guide), or a script run with Claude Code.

5. The document extraction agent

It reads unstructured documents (invoices, quotes, meeting notes) and extracts the data into a spreadsheet. No per-supplier rules are needed: the model understands different layouts. To build it: n8n or Make, with a watched folder as input and a spreadsheet as output.

6. The annotated monthly report agent

It pulls together figures from several sources, spots notable changes and writes a first commentary. The in-depth analysis stays human, but the collection and formatting are done. To build it: a no-code platform that fetches the exports, then an AI step that writes the commentary.

7. The internal research agent

It answers your teams' questions using only company documents, and cites its sources. To build it: a Claude project or a Gem with your documents attached, or a Copilot Chat agent if your company runs on Microsoft 365.

Other use cases, closer to automation than to agents, are gathered in my 10 AI automation examples.

Which free AI agent builder should you use?

There are two families of tools, and the choice comes down to a single question: should the agent act when you talk to it, or trigger on its own? Assistants (Gemini, Claude, Copilot) create conversational agents in a few minutes. Automation platforms (n8n, Make, Zapier) create agents that run without you, at a set time or when an email arrives.

Sources: vendors' pricing pages and help centers, checked on October 7, 2026. These offers change often: check them on the day you get started. With self-hosted n8n or Dify, the AI model itself is still paid per use, through the API key you provide.

My advice for a first try: a Gem or a Claude project if the agent works when you ask it to, n8n if you want it to run on its own and someone on the team knows how to install a service. To get started with n8n, follow my n8n tutorial, which also covers the free version.

How to build an AI agent: the 6-step method

The method is the same whatever the tool. The first four steps happen before you open any platform, and that is where success is decided.

Step 1: pick a precise, recurring task

A good first agent ticks five boxes:

• Recurring: the task comes back at least every week.
• Verifiable: a simple rule tells you whether the result is right.
• Low stakes if it fails: a missed monitoring report is easy to fix, an invoice sent to the wrong customer much less so.
• Already described: you can write down the steps you follow today.
• Accessible: the tools involved have an API or a connector.

If you cannot describe the task in writing, the agent will not guess it. That is the first cause of failure, well ahead of anything technical.

Step 2: write the agent's instructions

The instructions (the "system prompt") decide almost everything. They fit in five sections: the mission, the method, the output format, the success criterion and the rules. Here are the instructions of a monitoring agent:

MISSION
You are the competitive intelligence agent of [company], which sells [offer] to [audience]. Every morning, you spot what has changed at our competitors: [list of 3 to 5 competitors].

METHOD
1. Read the sources provided (websites, RSS feeds, pricing pages).
2. Keep only what is new since yesterday: offer, pricing, hiring, partnerships, major content.
3. For each new item, check that it actually appears in the source. If not, drop it.

FORMAT
A summary of 10 lines max: one line per new item, with the source link. If nothing has changed, write "Nothing notable today".

SUCCESS CRITERION
Every line cites a verifiable, dated source.

RULES
Never make up information. Never contact anyone. If a source is unreachable, flag it at the end of the summary.

The generator below assembles these sections from your answers and tells you which free tool fits best.

Step 3: list the tools and sources

For each step of the method, write down the tool the agent needs and check that access exists: a connector in the platform, an API, or an MCP server. Then sort every action into two categories: actions that read (no risk) and actions that write, send or modify (to be supervised). A first agent that only reads and drafts is already very useful.

Step 4: choose the building tool

• The agent works when you talk to it and reads documents: a Gem in Gemini, a Claude project or a Copilot Chat agent. Paste the instructions, add the files, and it is ready.
• The agent must trigger on its own (schedule, email, form) or write to your tools: n8n, Make or Zapier Agents.
• You can code, or the process is critical: a program written with an agent library (Claude Agent SDK, OpenAI Agents SDK, CrewAI), version-controlled and tested.

Step 5: test on ten real cases

Take ten past cases, including the three trickiest, and have the agent handle them. Compare with what you would have done. Every gap gives you a rule to add to the instructions. Then run the agent in observation mode for one to two weeks: it proposes, you decide, nothing goes out without you.

Step 6: set the guardrails and go live

Before letting the agent work unsupervised: human approval on every irreversible action, a cap on the number of steps per run to avoid loops, a log of every action, and an alert when the agent fails. Then add edge cases as they come up rather than trying to plan for everything up front.

Example: building a free monitoring AI agent, step by step

The same agent, built two ways. The first takes a few minutes, the second makes it autonomous.

On-demand version, with a Gemini Gem

1. Open gemini.google.com on a computer (Gems are created on the web) and create a new Gem.
2. Give it a name, then paste the instructions from step 2 into the instructions field.
3. Add as reference files the list of competitors and, if you have it, your latest monitoring note.
4. Test with a real request ("run today's monitoring") and fix the instructions after every disappointing answer.

The limit: the Gem does nothing until you write to it. It is a specialized assistant, not yet an autonomous agent. You can do the same thing in a Claude project.

Automatic version, with n8n

1. Schedule Trigger: runs every weekday at 8 a.m.
2. RSS Read (or HTTP Request): fetches the feeds and pages to monitor.
3. AI Agent: the node receives the instructions from step 2 and a model (Claude, GPT, Gemini or Mistral, through your API key).
4. Gmail or Slack: sends the summary to the team. Here, sending is safe, because it goes to you and not to a customer.

The n8n Community edition is free; you pay for the server and the model calls. You can build the same chain in Make with a scheduled scenario.

Building an AI agent with or without code: three paths

No code: assistants

Gems, Claude projects, Copilot Chat agents. Instructions, reference files, sometimes connectors. Best for: personal or team use, on demand. Limit: no automatic triggering.

No-code: automation platforms

n8n, Make, Zapier. A trigger, visual steps, an agent block in the middle. Best for: a non-technical team that wants to stay in control. Limit: complex logic quickly becomes unreadable, and debugging is painful.

With code: custom-built

This is the path I favor for anything that goes into production: a program written for the need, version-controlled, tested, with its own guardrails. With a coding assistant like Claude Code (see how to install Claude Code), writing that program takes a few days rather than a few weeks. Limit: someone has to keep it running.

How much does an AI agent cost?

Building an AI agent can cost nothing at first: the free plans in the table are enough for a first agent. The real costs show up later.

• Running costs. An agent makes several round trips to the model on every run. For an agent that runs every day on an average task, the order of magnitude is a few euros to a few tens of euros per month in model calls. The choice of model (a fast model is enough to sort or extract) and the amount of context you send can make that cost vary tenfold.
• Building. For a well-defined process that is already written down, count two to five days of work. For a fuzzy process, or one that touches several pieces of software, most of the time goes into clarifying the process, not into the technical side.
• Maintenance. Interfaces change, formats evolve, edge cases appear. Set aside time every quarter, or the agent quietly degrades.

The limits of an AI agent and the precautions to take

• Plausible errors. An agent can make up information with total confidence. Hence the verifiable success criterion and the instruction to cite its sources.
• Confidentiality. An agent that reads your files sends their content to a model. Check the terms of service, and favor business plans that exclude your data from model training.
• Traceability. Every action must leave a trace: if something goes wrong, you need to be able to say what happened and why.
• Decisions about people. Rejecting an application or screening candidates fully automatically falls under specific rules (in the EU, the GDPR regulates decisions based solely on automated processing). Keep a human in the decision.
• Dependence. An agent at the heart of a process ties you to a platform and a model. Document it and keep your data exportable.

AI agent or classic automation: how to decide

Many companies build an agent where three conditions would have been enough, and pay more for a less predictable result. My rule comes down to one question: are the steps always the same?

Most of the time, the best architecture combines both: a classic automation orchestrates the trigger and the stable steps, and only calls the AI on the part that requires understanding. To map out a complete process and decide step by step, follow my business process automation method. And if you would rather delegate the build, that is exactly what I do in an AI automation engagement.

Written by

Anthony Courtin

SEO, GEO & AI automation consultant

I help SMBs and independent professionals with their visibility (Google, ChatGPT, Perplexity) and with the real-world use of AI at work. I work with companies on site to train teams and rethink processes. Everything I write here, I use in client work.

Follow me on LinkedIn My YouTube channel (in French)