ajsai47/autocode

Seven specialized AI agents execute a full development pipeline—Scout → Architect → Builder → Tester → Reviewer → Ship—pulling work from a unified queue fed by GitHub Issues, coverage gaps, backlog tasks, and proactive codebase discovery. Each agent has strict read/write boundaries and runs at progressively harder difficulty levels, starting with simple test coverage and graduating to feature work. Uses a manifest-driven architecture for reproducibility, isolated git worktrees for safety, and persists learned patterns across cycles via a local knowledge graph that scores fix success by recency and outcome.