HarnessOpt-Bench tests LLMs on agent harness optimization

Researchers introduced HarnessOpt-Bench, a benchmark for evaluating how well LLMs perform end-to-end harness optimization for agentic systems. In the protocol, an LLM paired with a coding harness edits a target agent's seed harness using graded evaluation feedback and a fixed target-evaluation budget, then submits a final candidate scored by normalized gain on an inaccessible held-out test partition. The benchmark uses a trusted execution environment to enforce evaluation boundaries, meter target-agent resource use and preserve candidate versions for audit.