Claude+small-model harness cuts 1,000 AI decisions from $605 to $0.17
jamestagg · x · 2026-09-30
polydao shares an agent harness skill claiming 1,000 decisions cost $0.17 instead of $605 on a cold frontier model. The core idea: "Claude proposes, Jev answers, code decides," with a receipt logged at every step.
The six-step loop:
- propose: Claude writes plans and patches but decides nothing
- filter: code drops routes the machine can't run before Jev sees them
- answer: Jev picks probabilistically from code's menu, or abstains
- re-check: code validates picks against live state; stale results go to fallback
- act: tools run through the normal approval path
- receipt: every step is logged and replayable against new policies
Benchmarks: 0.2s median per call, $0.0002 per desktop step, 49/49 WebMCP tasks passed. Key insight: "I don't know" is a real answer — at 38% confidence Jev falls back to a pre-written action instead of forcing an expensive model to guess.
More from coding & agent
- ccsession adds Snowflake CoCo CLI support for fuzzy-resuming agent sessions — 4310sy · 2026-09-30
- shadcn creator: copy the components freely, but don't clone the docs site — shadcn · 2026-09-30
- 61 days of Claude Code transcripts: hooks reported success but only 11% of context reached the model — JhouHate · 2026-09-30
- Open-source Codex skill scrapes Xiaohongshu related keywords, scoring 93 terms per seed word — vista8 · 2026-09-30
- Constella agent drives Blender via MCP: rough edges, high upside — sidahuj · 2026-09-30
- Law firm AI working group tests multiple models — Claude keeps outperforming — jkubicki · 2026-09-30