Execution Engine
Execute, review, learn, ship. One command orchestrates parallel subagents through the full implementation cycle with automatic knowledge capture.
Main Agent
Orchestrates and dispatches subagents
Each subagent follows this pipeline independently. The main agent orchestrates, dispatches, and verifies.
Inject relevant knowledge from previous work
Follow well-planned tickets in beads
fix loop
(max 3x)
Architecture, security, performance, debug leftovers
Fix issues found in review, then re-review
Log knowledge learned during session
Run tests, lint, close beads ticket
Pass an epic ID and Lavra auto-routes to parallel execution. Beads are organized into dependency waves -- each wave runs concurrently, respecting file ownership boundaries.
Wave 1
Infrastructure
No dependencies. Runs first.
Wave 2
4 components in parallel
File ownership prevents conflicts.
Wave 3
Assembly + polish
Depends on all Wave 2 beads.
One command.
Entire feature shipped.
Recall. Implement. Review. Learn. In parallel when possible.