Researchers evolve specialized agent harnesses instead of one global harness
An Oct 20 arXiv paper introduces OO-Harness, which evolves a hierarchy of specialized harness subclasses and routes each task to one at test time.
A new preprint, Object-Oriented Harness (arXiv, October 20, 2026), argues that one global agent harness handles mixed task types poorly. Its fix is to evolve a family of specialized harnesses and route each task to one of them. If you run one mega-prompt agent across very different jobs, this paper tests the alternative.
Per the abstract, during training an "evolution agent" uses execution records to build specialized harnesses. Each is a subclass of the base harness or of another specialized one. At test time, a router uses lessons gathered in training to pick a harness per task. Across three long-horizon benchmarks covering business workflows, workspace tasks and tool use, the authors report the best results among compared methods, with an absolute gain of up to 15% over the base.
In a thread on X, @omarsar0 attributes the paper to Google Cloud AI Research. He says it lifts AutomationBench from 61.6% to 77.4% with the same Gemini 3.5 Flash executor. He also says spending extra inference on best-of-K sampling or self-refinement did not reliably help, and that subclassing beat the same evolved code copied into flat classes.
An independent paper, Mixture of Self-Improving Branches, also routes tasks to specialized harnesses. It reports relative gains over Meta-Harness of 11.6% on Terminal-Bench 2.0 and 3.8% on SWE-bench Lite.
This is a research result, not a tool. The abstract doesn't say whether code is available. The sources don't give the cost of the evolution step or how often the router picks the wrong harness.
If you want to experiment now, the Meta-Harness onboarding guide gives two pieces of advice:
- Keep final test data out of the optimizer's reach.
- Run a few iterations to check proposal quality and cost before committing a larger budget.