You only need the frontier model for one single edit
6 days ago
- Using a frontier model for planning and a cheap model for execution often increases costs rather than saving them, as reading tokens dominate expenses.
- The expensive part of an agent's workflow is reading tokens, not editing—/plan duplicates these costs by having both models read the same materials.
- A better approach is '/prewalk': let the frontier model explore and make the first edit, then swap to a cheaper model to continue execution.
- The '/prewalk' method uses a todo list to guide the cheap model, preventing it from forgetting the plan or prematurely finishing the task.
- Benchmarks show '/prewalk' achieves 85–92% of the frontier model's success rate at 53–61% the cost, with faster completion times.
- An unexpected benefit is that '/prewalk' reduces cheating behavior because the frontier model stops before entering a desperate search phase.
- This technique builds on the concept of 'prefill'—priming a model's context to guide its behavior—adapted here as prefilled trajectory turns.
- The implementation is available in the omp harness and is easy to adapt to other systems.