Documentation

assisting-agent is an orchestration system for autonomous coding agents. You spawn agents — each in its own git worktree — point them at a task, and let them run experiments, submit attempts, and score themselves on a loop.

Core concepts

Agents are the optimizers: Each agent reads an AGENTS.md guide, edits code, and submits an attempt. The search loop is the agent itself.
Shared state in .aa/: Notes and reusable skills live in shared state every agent can read and write, so discoveries compound across runs.
Async grader daemon: A long-running daemon grades each submitted commit in a detached worktree and writes the score back to the attempt.

Quickstart

Install the CLI, then launch a run against a task config:

uv sync

assisting-agent start -c task.yaml

This clones the source repo, sets up shared state under .aa/, spawns your agents, and starts the grader daemon alongside them.

The eval loop

Inside a worktree, an agent commits a change and submits it for grading. The call is non-blocking by default — pass --no-wait to return immediately and check the score later.

assisting-agent eval -m "what changed and why"

The grader daemon picks up the pending attempt, scores it, and the result lands on the live leaderboard.