Google paper: gated self-improving agent harnesses cut token cost 10x and transfer across models
omarsar0 · x · 2026-11-18
Swerling observability is out; harness engineering is in. A new Google paper, "Harnesses That Improve Themselves," proposes a framework where an optimizer rewrites the agent runtime around a frozen model.
Key results on SWE-bench Lite (frozen Gemini 3.8 Flash):
- Rewriting a prompt rule file actually lowered held-out accuracy from 5.0% to 2.5%.
- Full harness rewrite: 42.5% accuracy at 451.6k tokens per task.
- Gated edits to one module at a time: 40.0% at just 46.7k tokens per task (10x cheaper).
The gated approach splits the harness into four editable components — context, control, tools, verification — patches one per round, validates on held-out tasks under a token cap, and deploys only if it passes.
It also transferred to Gemini 4 Argon with no new search (37.5% vs 30.0% for full rewrite). Caveat: only 40 held-out issues, so treat numbers as directional. The paper catalogs failure modes of self-modification, including prompt bloat, benchmark gaming, and evaluator tampering.
More from coding & agent
- OTel Node.js SDK lands 1-line change: Object.keys swap gives 2.4x speedup — DanielLockyer · 2026-11-18
- Academic Research Skills for Claude Code hits 57.5k GitHub stars — adnan_hashmi · 2026-11-18
- Building your own agent harness: why banks and hospitals skip off-the-shelf coding agents — TejasKumar_ · 2026-11-18
- Replan projects via MCP: a two-step prompt pattern for Codex and Claude Code — sfoooooooooooooooooo · 2026-11-18
- CallForMe plugin lets ChatGPT phone US businesses for quotes and bookings — jakemarsh · 2026-11-18
- Probe Volumes lands in PlayCanvas engine with GPU re-baking and day/night setups — willeastcott · 2026-11-18