Google's OO-Harness uses object-oriented design for self-evolving agents, lifting AutomationBench from 61.6% to 77.4%
dair_ai · x · 2026-10-21
Google Cloud AI Research's new paper introduces OO-Harness, a self-evolving agent harness built on object-oriented design. With the same Gemini 3.5 Flash executor, it raises AutomationBench from 61.6% to 77.4% and beats Meta-Harness on all three benchmarks tested.
Instead of evolving one global harness for every task, an evolution agent builds a class hierarchy of specialized harnesses from execution records:
- Subclasses inherit the base harness and override only the methods a task cluster needs
- Design lessons record which changes helped and which failed
- A router reads the task and dispatches it to the right specialized harness
The authors' takeaway: build specialized harnesses rather than chasing one universal setup.
More from coding & agent
- EAU: Co-evolving user simulator, environment, and RL setup for training LLM agents — UAlberta · 2026-10-21
- Meta's Self-Credit: Dense Rewards from Sparse RLVR via Intra-group Rescaling — meta · 2026-10-21
- Tamper-proof agent logs still can't prove the agent did the job: who reads the record back? — ZealousidealAct9661 · 2026-10-21
- One prompt: Claude builds an animated mech spider in Blender via MCP, first try — sidahuj · 2026-10-21
- OpenAI Details Coordinated Disclosure of Critical Vulnerabilities Found by Agents — deliciousmark12 · 2026-10-21
- Dev: coding agents make every idea 10 minutes away — and that's the problem — feregri_no · 2026-10-21