harness

Multi-agent coding sprints are orchestrated by dispatching initializer, planner, generator, evaluator, coordinator, and releaser agents to execute supervised or continuous development rounds. This approach turns short prompts into repeatable sprinted workflows that complete features one at a time.

Feature tracking prevents lost work by maintaining a machine-readable features.json ledger and applying auto-commit protocols for per-round artifacts. This ensures measurable progress and stateful round management across long-running application harnesses without context loss.

Automated sprints enforce code validation through an authenticity gate, quantified evaluator scoring, and artifact validation. These blocking gates ensure features actually pass rigorous evaluation before being flipped to complete, preventing incomplete work from accumulating.

Context limits are managed through context reset handoffs between agents during multi-agent development sessions. This process clears stale state between rounds, while error recovery mechanisms handle retries and explicit stop or pause rules during continuous runs.

Scaffold a multi-agent sprint environment by running the start command to create the .harness directory, which establishes the project structure. Then execute the session command to run a single supervised sprint round for incremental feature implementation.

Sprints accumulate incomplete features when the authenticity gate and quantified evaluator scoring are bypassed or misconfigured. Enforcing rigorous separation of generation and evaluation ensures features pass validation before advancing, keeping progress measurable.