Google paper: object-oriented specialized agent harnesses hit 77.41% vs 61.62% baseline
rohanpaul_ai · x · 2026-10-22
A Google Cloud AI Research / UIUC paper introduces OO-Harness, a framework that evolves agent harnesses (the prompts, tools, and control code around a model) using object-oriented design.
- Most tuning methods evolve one global harness, but a fix for one task type often hurts another
- OO-Harness implements harnesses as code classes: specialist harnesses inherit from a parent and change only what their task cluster needs; a router picks a specialist or keeps the base result at test time
- On AutomationBench business workflows it scores 77.41% vs 61.62% for the base harness and 63.92% for the global Meta-Harness method
- On unfamiliar apps the router keeps the base harness, holding a 59.18% pass rate while Meta-Harness's global harness falls to 44.90%
The paper recommends keeping the base harness as fallback and routing to specialists only when a run shows a known problem. Across three long-horizon benchmarks it delivers up to 15.79% absolute improvement over the base.
Related event: Google's OO-Harness Paper Boosts Self-Evolving Agent Performance to 77.4%(3 posts)→
More from coding & agent
- OpenClaw opens its team environment to community contributors via Focus Weeks — RichardSocher · 2026-10-22
- ESA's cat keyboard incident is a warning for leaving AI agents unattended — vincent_koc · 2026-10-22
- Defense must run at machine speed: agents patch, engineers judge — saranormous · 2026-10-22
- TraceForge synthesizes reasoning traces from answers, beating Claude's own CoT by +3.7/+7.1 — IndependentResearcher121 · 2026-10-22
- Internet Archive fends off PHP attacks overnight, using Claude to patch old code — kingslyj · 2026-10-22
- Combining Ghidra, AI Reverse Engineering and WebGPU Sparks an "Absolute Renaissance" — thursdai_pod · 2026-10-22