Google's OO-Harness lifts agent success from 61.6% to 77.4% without changing the model
alex_verem · x · 2026-10-24
Google Cloud AI Research and UIUC evolved agent harnesses with object-oriented design (arXiv:2610.28380) and gained nearly 16 points without touching the model.
How it works
- The harness is code around the model deciding prompts, tools, and completion checks
- Claude Opus 4.8 acted as the evolution agent, writing specialized harnesses that inherit from a base and change only what one task group needs
- A router picks the harness per task; it kept the original setup on 134 of 300 tasks
Key numbers
- Business workflows: 61.62% → 77.41%, fully solved 25.44% → 37.22%
- Office file tasks: 25.45% → 39.48%
- On unseen apps it falls back to base and keeps 59.18%, while the compared method drops to 44.90%
- Search done in 29h vs 109h for the older approach
Counterintuitive: best-of-5 sampling scored 60.24% on business workflows, below a single run.
Takeaway: if your agent keeps failing, check the code around it before paying for a bigger model.
More from coding & agent
- Microsoft launches Agentic DevOps Engineer Expert certification via Exam AZ-401 — adnan_hashmi · 2026-10-25
- PL researcher admits Opus 5.5 is already a better programmer than him — tweetsatpreet · 2026-10-25
- Free LLM Dojo launches: 50+ interactive labs and 490 AI-engineering interview questions — sharpeye_wnl · 2026-10-25
- mlx-decision runs decision models locally on Apple Silicon, up to 6x faster than PyTorch — armbues · 2026-10-25
- How do teams share memory across multiple AI agents? — belhelha · 2026-10-25
- Vibe Monitor ships after 41 beta iterations: local stdio MCP adds architecture checkpoints to coding agents — Either-Alternative32 · 2026-10-25