Decide what to work on, dispatch agents in parallel, review output, iterate. Each agent runs in its own Docker container — no permission prompts, no input needed.