Create an AI agent

The simplest way to create an AI agent today is not to program it from scratch: use a tool that already comes with the agent built in, like Claude Code or OpenClaw, and give it instructions, access to the tools it needs and clear limits. Creating an agent is, above all, deciding what it has to do and what it's allowed to touch. The technical part, these days, is the least of it.

Every agent has the same ingredients: an AI model that thinks, instructions that explain the task and how to decide, some tools (connections to your email, your files or your programs, often through MCP, a standard for connecting AI to other apps) and some permissions. If you're a developer, there are libraries for building them to measure, but you don't have to start there. What's confusing are the "build your agent in five minutes" videos: the agent gets built quickly; the hard part is getting it to do the task well without making a mess.

There isn't a blog article carrying this question yet. In the meantime, my practical recommendation: start with a small, reversible task, in a test folder, with no access to your email, your clients or anything that costs money. Watch what it does, read what it proposes before accepting it and only give it more permissions once it has shown it can handle the previous ones.