Short answer: you do not start with the tool, you start with the job description. An AI agent is a role, a scope, its refusals and its escalation rule — written once, held over time. The free version of Claude, ChatGPT or Gemini is enough to start. Zero code, zero subscription to a third-party tool.
Most searches start from the tool: “create an AI agent with n8n”, “with ChatGPT”, “with Make”… The tool matters less than you think. A well-defined agent works in any of the general-purpose AIs; a badly defined agent drifts in all of them. What makes the difference: the structure you give it before the first conversation.
Start from the real work: which recurring task eats hours every week — follow-ups, quotes, document sorting, posts? That task becomes the agent. One agent = one mission, not a Swiss army knife expected to do everything.
Four lines are enough: what the agent does, what it never does, its style, its output format. Refusals matter as much as missions: they are what prevents drift.
Whatever falls outside the scope goes back to the human — explicitly, in writing. An agent without an escalation rule makes things up, and you end up correcting in a loop.
One complete task, with the real files. Where it drifts, tighten the job description — not the prompt of the day.
By default, the agent prepares and every output waits for validation. After ten validations in a row with no edits, delegation is earned — and can be revoked with one word. You delegate a scope, never the business.
Claude, ChatGPT or Gemini: any of the three is enough, including the free version to start. Searches like “create an AI agent with n8n” or “with Make” lead to powerful automation tools — useful when dozens of apps have to talk to each other, unnecessary for a first agent. The right sequence: an agent that works in a general-purpose AI, then automation only if the need proves it.
On the data side, two questions deserve asking before connecting anything: where the information pasted into the conversation goes (the privacy scope of the chosen AI and its settings), and what stays human. The safest setup is still the one that runs with a written scope.
The fastest way to know which first agent to build is the Evol-AI Scan: a free 30–45 min diagnosis that stress-tests the idea, no detours, and gives a clear verdict plus an exportable plan. The free version of Claude is enough to run it.
The human pilots. The agent executes
Launch the free Scan →once the agent works on its own, the team comes next: how to build an AI agent team · or build directly with TEAM BUILDER · Pro and Premium