OO-Harness is claimed to lead compared methods on three AI agent benchmarks

OO-Harness's authors describe agent harnesses as code that sets a language model's prompts, tools and control flow. Their system evolves specialized subclasses from execution records, with a router selecting a harness for each task. The authors report the best performance among compared methods on three long-horizon benchmarks. A post sharing the paper says its AutomationBench score rose from 61.6% to 77.4% using the same Gemini 3.5 Flash executor.