Unlazy makes AI work prove completion

Unlazy targets the biggest weakness in agentic coding: confident completion without verified completeness. Its strongest idea is making “done” an auditable state rather than a model-generated claim.

• –Gates store acceptance criteria, commands, expected outputs, and evidence directly in project files
• –The Depth Tree breaks large builds into fresh-context leaf tasks, reducing long-context drift
• –Parent re-verification limits the risks of agents self-certifying their own work
• –Claude Code users can add a Stop hook that blocks completion while gates remain unmet
• –Parallel execution can improve throughput, but increases token costs and coordination risk