agentbrain

Agent Brain is a portable set of files and rules that you add to a code repository to give AI coding agents better operating discipline. It does not replace or run the agent itself. Instead, it layers on top of whatever coding AI you already use, it explicitly supports Claude Code, Codex, Gemini CLI, Cursor, Copilot, Windsurf, OpenCode, OpenClaw, and Hermes, and gives that agent a structured process to follow.

The problem it addresses is that most AI coding agent failures are not technical errors but judgment errors: the agent builds the wrong thing, trusts outdated context, skips tests, accepts vague requirements, or ships without a rollback plan. Without structure, agents charge ahead based on chat momentum rather than evidence.

Agent Brain enforces three habits. First, plan before build, vague requests must pass through intake, research, challenge, and design before any implementation starts. Second, verify before trust, tests, logs, diffs, screenshots, and approvals count as proof, while confident summaries from the agent are treated as claims until backed by evidence. Third, learn only from evidence, repeatable successful workflows get distilled into small, neutral skills rather than absorbed carelessly.

The system supplies a state machine, command specifications, skill definitions, schemas for artifacts, evaluation gates, and handoff contracts. A request flows through this pipeline: raw request, state, command, skill, artifact, evidence, handoff. Nothing proceeds to the next stage without the required artifact from the previous one.

It is documentation-first and written in Python 3.11, released under an open source license visible in the repository.