A developer cut frontier-model usage 75% by giving the harness five roles

Fryxell runs three coding-agent TUIs, Cursor, Claude, and Pi, sharing one set of skills and one AGENTS.md, and defaults to the cheaper deepseek-v4-flash-0731 for maintenance work, reaching for frontier models only when exploring a large feature or refactor. He built that split into a five-stage pipeline: explore, plan into an explicit task DAG, worker, critic, and a fifth stage he added himself, promoter, a forced step for writing up and communicating finished work, which he says he needed because he tends to ship and move on. He also adopted Can Bölük's "prewalk" technique, using a frontier model for planning and the first task, then handing off to the cheap model once the pattern is set. This, he writes, "has dropped my usage of Fable in even my most intense contexts by 75%," a change he says "feels sane, something akin to AI soak." He separately describes routing his own editor configuration through the same harness as "harness as jig."

Why it matters: The role split gives readers a specific setup to test: use a frontier model for planning and critique, a cheaper model for execution, and require a write-up at the end. The 75% figure comes from one person's own use and has not been independently measured, so it is evidence from one setup rather than a benchmark.