A solo developer documented a four-role harness that cut his frontier model usage 75%.

Scott Fryxell runs planner, worker, critic and promoter as separate roles, with a frontier model doing the "prewalk" planning and refinement while commodity models such as Deepseek-v4-flash execute, on two $20/month subscriptions plus occasional Pi access (scott-fryxell.github.io). He runs three TUIs (Cursor, Claude, Pi) over one shared AGENTS.md with skills, extensions and artifacts directories treated as a reusable jig. He isolates each role after finding that a single prompt that plans, executes and critiques confuses its own objectives. No benchmark, one practitioner, but the cost structure is concrete and the role-isolation finding matches what I've hit in my own loops.