Working with AI agents

How to brief, supervise, and verify AI agents working on recomp projects.

This section is for people using AI agents on recomp projects.

Do not assume an agent will discover these rules by crawling the site. Treat these pages as a briefing packet: give the agent the rules, point it at the local repository instructions, and make it prove its work.

• What to tell an AI agent before it touches a recomp project. The human-facing overview and recommended agent setup.
• Rules to give an AI agent. The rules an agent should follow unless the local repository says something stricter.
• How to check AI work. What proof counts for builds, coverage, co-simulation, screenshots, TCP checks, and releases.
• How AI breaks recomp projects. The quiet failure modes to look for when reviewing AI-generated recomp work.
• Debug surfaces agents can use. TCP servers, traces, JSON output, screenshots, input commands, and other surfaces an agent can query.
• Making AI contributions reviewable. Commit rules, handoff expectations, and what never belongs in a change.
• When an agent cannot run the game. What an agent can still check without the game, display, or oracle.
• After an agent finishes, expect a recap that names the files changed, checks run, commit, and published URL.