Most developers have by now asked an AI agent to build something. Sometimes the result is impressive, sometimes it is wrong in a way that takes longer to untangle than writing the code yourself. The difference is rarely the model. It is how the agent is steered.
What an AI agent does
An AI assistant answers questions in a chat. An AI agent works on its own in your project: it reads the code, runs commands, writes and changes files, runs the tests and tries again when something fails. Claude Code, GitHub Copilot and Codex are examples of tools that work this way.
In other words it is no longer about getting code suggestions. The agent does the whole task and you review the result. That is where both the gain and the risk are.
Where agents are strong
Agents are fast at work with a clear answer. They write tests, move code, swap one library for another, rewrite scripts and document what already exists. Work that used to take a day often takes an hour.
Where they need steering
The agent guesses wherever you are vague and it guesses with great confidence. It picks its own solution for something you took for granted, misses a security requirement that was never written down or builds more than you asked for.
On top of that, whoever wrote the code sees the same things as when they wrote it. An agent reviewing its own code rarely finds what it missed the first time.
Four things that make the difference
- A spec before the code, with acceptance criteria that can be tested. The agent stops guessing when it knows what to build.
- A constitution with the rules that always apply, for example that tests are written first and that no secrets end up in the code.
- A separate reviewer. Let another agent with its own brief check the change against the spec.
- A person who decides on what goes to production, on personal data and on trade-offs between security and simplicity.
How we work
Activator.se is built with AI agents this way. The project has a constitution with eleven principles, just over 60 specs and more than 2,000 automated tests. Every change is reviewed by a code reviewing agent and a security reviewing agent. Nothing is deployed without a person approving it.
Honestly, what surprised us most was not the speed. It was that the questions came before the code instead of after.
Get started
The easiest first step is to try the workflow on a small feature. In the article GitHub Spec Kit: how to get started we go through installation and the commands. The whole way of working with examples is in the guide Build it right with AI agents.
