Scott Fryxell’s thesis, stated plainly: the harness is the thing — the fulcrum where your expectations meet the LLM’s capabilities. Eighteen months of tab-completion → agentic coding → managing agents with a harness, and the constant conversation about how we solve problems is itself the engine of progress (his graybeard’s “Moore’s law also applies to software”).

The rig: commodified models, unified experience

Two subscriptions (Cursor, Claude) plus Pi as needed — and all three share the same skills and AGENTS.md, so the models are interchangeable. “There is no magic sauce… I have zero anxiety about the transition from Cursor to Codex.” Cost-wise he leans on deepseek-v4-flash-0731 for maintenance and simple tasks, dipping into the frontier (Fable) only for serious features and refactors — and has cut even that by 75% using the prewalk technique (frontier for planning + first task, then hand off once the pattern is set) combined with the planner/worker/critic split: a single prompt that plans, executes, and critiques itself confuses its own objectives, so each role gets isolated.

The arc: explore → plan → worker → critic → promoter

  • Exploration leads to a plan formalized as an explicit DAG task list
  • A worker implements the DAG one node at a time
  • The critic simplifies and questions what was built — often pushing back into another worker pass
  • The promoter is the reminder that a job isn’t complete until it’s communicated — added deliberately to shore up a ship-and-move-on weakness, and run through the frontier because “promotion is subtle and easily borked, and nobody likes a critic”

The result: 75% less frontier spend even in his most intense contexts — “AI soak” on two twenty-dollar plans.

The harness as jig

His working rig is brayness: nvim config mapped into the work directory so the LLM knows what file is open and can edit it with him — harness as jig, keeping him disciplined about vim motions. Everything is first-class at the root — AGENTS.md, AGENTS.local.md, bin/, prompts/, plans/, skills/, extensions/, artifacts/, work/ — with a claude() shell wrapper that auto-loads the nearest AGENTS.md, and a skill that syncs the harness back to the repo. Auditability is the rule: TUIs are instructed to keep artifacts inside artifacts/, which is why the root is deliberately git-less.

Product tie-in: his camera app (realness.online — photo → vector graphic) is exposed to the harness via a poster-driver headless-Chrome script the LLM wrote once and keeps in the harness, so it runs a billion times without burning tokens.

What he learned

  • Skills were initially too prescriptive: “there is a line past which you are burning tokens mansplaining to clankers” — and he needs a more empirical approach to confirming the impact of harness changes
  • The government’s Fable ban turned a hundred thousand developers into model-diversifiers overnight; Pi went from a toy to the most important piece of the rig
  • Leaving the editor for the terminal more and more, trusting agents with the code — and his output is at its highest level and tempo

The through-line for anyone running an agent setup: skills, extensions, and AGENTS.md as first-class citizens at the root, the models themselves commoditized underneath — and the harness carrying the parts that repeat, so the frontier only touches the parts that don’t.