If you’re about to let a coding agent (an AI tool that writes code changes) loose on your code, this page is for you. The code in a run comes from agents. This page is about how they’re given the job, what they can and can’t reach while they work, and how Evolve treats what they say about their own work.
Any agent that can work by itself
Evolve asks an agent to make the change. It has run three well-known coding agents, Claude Code, Codex and opencode, from start to finish. Any tool that can take instructions and edit files without a person at the keyboard can be plugged in. There’s also a scripted stand-in for demos, which says plainly that it isn’t an AI model.
A sealed workspace for each attempt
Each attempt gets:
- its own copy of the code, so attempts can’t get in each other’s way;
- a sandbox: a sealed-off space on your machine that hides your home directory, Evolve’s own data and the key Evolve signs its results with;
- only the logins that agent needs, copied in for the attempt and removed afterwards;
- the tools your project is built with (Go, Node, Python or Rust, depending on its programming language), so it can run your tests while it works;
- network access only if your policy (the rules file in your project) allows it. Agents usually need it, to reach their model: the AI they talk to over the internet.
When the agent finishes, Evolve itself works out exactly which files changed and turns the result into a commit, a saved snapshot of the code in Git.
What an agent says about its work
Agents report on their own work: “all tests pass”, “prices unchanged”. Evolve records those reports next to the attempt, marked as untrusted, and they play no part in the decision. The decision comes from the evidence the checks produce: the results of tests Evolve runs itself.
More than one attempt
A run can ask for several attempts: a few different approaches from one agent, or one each from different agents. More attempts mean more chances that one of them is both fast and right.
Next
Checks and evidence: how an attempt is judged.