I built a multi-agent loop where an adversarial Claude reviewer reads your actual codebase before approving plans

Short summary

The author built agent-plan-review-loop, an open-source system using Claude to adversarially review implementation plans by reading the actual codebase rather than relying on author context. The system uses tiered task classification to optimize model costs, isolates implementations in Git worktrees, and includes human approval gates for business decisions. This architecture moves feedback from conversational agreement to independent, skeptical review based on artifact files.

• •Multi-agent system uses adversarial Claude reviewer to catch plan flaws by analyzing actual codebase and artifacts, not conversation context
• •Task-based cost optimization: T0/T1 use Sonnet, T2 uses Opus, reducing expensive reasoning calls while maintaining quality
• •Isolated Git worktrees, human decision gates, and file-based memory enable safe autonomous workflow with human oversight

Generated with AI, which can make mistakes.