How to build an AI agent, no code

An AI agent is an AI with one job, a personality, rules it always follows and skills for the tasks it repeats. Here’s how to build one in eight steps, with a worked example. Updated 24 October 2026.

Skip ahead: use the free builder

AI agent vs chatbot

A chatbot answers what you ask and forgets. An agent is set up for a job: it knows its role, follows a routine, keeps to rules and can work through several steps (find the overdue invoices, draft the reminders, ask you to approve them). You don’t need a new app for this. ChatGPT, Claude, Gemini and the rest can all run agents once you give them the right files.

The four parts of every AI agent

• 🧠 Instructions: who it is, its job and how it works.
• 🛡️ Rules: what it must never do without asking.
• 🧰 Skills: step-by-step know-how for repeated tasks (SKILL.md files).
• 🔌 Tools: what it can reach (email, calendar, files), added carefully, last.

How to build an AI agent in 8 steps

1. Pick one job. The most common mistake is an agent that does “everything”. Choose one job you repeat every week: chasing invoices, answering enquiries, writing a morning briefing, booking appointments. One job means clear instructions and results you can check.
2. Give it a name and a personality. A name makes the agent easy to talk to and easy to tell apart from your others. Then decide how it should sound: formal or casual, brief or detailed, cautious or bold. Write it down in a sentence or two (“Calm, brief and polite; never pushy”).
3. Write its instructions. This is the agent’s core file: who it is, what its job is, what it should do step by step, and what a good result looks like. Platforms call it different things (project instructions in ChatGPT and Claude, a SOUL.md in OpenClaw and Hermes, CLAUDE.md or AGENTS.md for coding agents), but the content is the same.
4. Add rules it must never break. Rules are what make an agent safe to use. The essentials: draft, never send without approval; ask before spending money or deleting anything; never invent facts or figures; keep personal data private. Put them in the instructions and repeat the important ones in each skill.
5. Turn repeated tasks into skills. A skill is a short SKILL.md file for one task the agent does again and again, such as “draft the next invoice reminder”. The AI loads it only when needed, so the task is done the same way every time. Most platforms now read the same SKILL.md format.
6. Choose where it lives and install it. Pick the AI you already use. ChatGPT and Claude use Projects plus uploaded skills; Gemini uses skills (Gems are being phased out); Claude Code and Codex read files in your project folder; OpenClaw and Hermes run your agent on your own computer and can talk to you on Telegram.
7. Test it before you trust it. Run five checks: ask it what its job is; give it a real task; ask it to do something it shouldn’t (like sending an email) and make sure it asks first; give it incomplete information and see if it asks rather than guesses; start a fresh chat and check it still knows who it is.
8. Connect tools slowly. Once it behaves, let it use tools one at a time (email drafts, calendar, files), always with approval switched on. Widen its permissions only when it has earned them.

Worked example: an invoice-chasing agent

# Chase, Invoice Chaser
You keep track of every unpaid invoice and draft reminders
that get firmer step by step: friendly nudge, reminder,
firm, final. You keep a running total of what is owed.

## Rules
- Draft only. Never send without approval.
- Never threaten, and never agree a payment plan yourself.
- If an amount or date is missing, ask. Never guess.

Where to run your agent

All of these read plain-text instructions and SKILL.md skills. Each has a step-by-step guide:

• 👉 ChatGPT setup guide
• 👉 Claude setup guide
• 👉 Claude Code setup guide
• 👉 Codex setup guide
• 👉 OpenClaw setup guide
• 👉 Hermes setup guide
• 👉 Grok Bot setup guide
• 👉 Gemini setup guide

Common mistakes

• ❌ Too many jobs. Split it into two agents instead.
• ❌ No rules. An agent without “ask first” rules will happily send a half-finished email.
• ❌ Testing in an old chat. Most platforms only load new instructions in a fresh chat.
• ❌ Connecting everything on day one. Add tools one at a time, with approval on.

Build one in five minutes instead

The KitMyAgent AI agent builder does steps 1 to 5 for you: pick a job, a look, a personality and the rules, and it writes the instructions, the skills and the setup files for 8 platforms. Need inspiration first? Browse AI agent examples or try the name generator.

Build your agent free

Questions