The idea: Claude acts as orchestrator: it plans tasks, breaks them into bounded units, and judges the output. Cursor agents (basic plan, almost unlimited and super fast) handle the actual code generation, which doesn't need frontier-level reasoning.

A guarded loop (PLAN -> BUILD -> JUDGE) enforces safety constraints between steps (no destructive ops, scoped file access, bounded iterations).

In practice, this brings my AI-assisted coding cost to ~$40/month while keeping the quality of having a strong model in the loop for architecture and review decisions.

The core tradeoff: you give up the simplicity of one model doing everything, but you gain cost control and a natural separation between "thinking" and "typing."

Repo: https://github.com/clementrog/claude-relais

Would love feedback on the orchestration approach, especially if others have tried similar multi-model setups and hit failure modes I haven't seen yet.