Harness — Multi-Agent Orchestration

Orchestrate complex tasks through Planning -> Generation -> Evaluation -> Retro. Fresh sub-agents per checkpoint prevent drift. Retro accumulates learning across tasks.

Recommended Workflow

Session 1 (Claude Code) → Brainstorm + Spec → Spec review with cross-model evaluator
    ↓ spec approved
Session 2 (Codex)       → Execute checkpoints → Evaluate → E2E → Full-verify → PR → Retro
    ↕ Claude CLI as review-loop peer (cross-model quality gate)

• Session 1: Claude Code for interactive discovery — brainstorming + multi-turn Q&A to establish requirements. After the user approves the brainstormed design, spec drafting and the Spec Evaluator review loop run autonomously (agent-to-agent until consensus). See references/planning-protocol.md "Post-Brainstorming Autonomy".
• Session 2: Codex for autonomous execution — implementation, evaluation, PR creation, retro. Claude CLI serves as cross-model reviewer via review-loop.
• Both hosts support both phases. The above is the recommended flow, not a hard constraint.

Prerequisites