EnvHarness adds 9 points to agent scores

Google Cloud AI Research’s EnvHarness reported up to a 9.0-point gain on held-out instances across five benchmarks, with 9.8% fewer execution steps, by wrapping static environments without changing their tasks or verifiers. The project page and the accompanying paper describe an open-source layer that lets an LLM designer target an agent’s diagnosed weaknesses. Builders should treat the environment as part of the training loop—not just the model or harness—and test whether adaptive feedback improves their own benchmark, extending the archive’s harness-cost analysis.