Did Codex Reset
GitHub

Harness-Zero raises macro task success to 44.3% without the specialized harness

elvis

A paper from Google and colleagues studies whether an agent harness can be distilled into a model. Harness-Zero keeps the optimized harness in training only. Because the optimized harness and the deployment harness use different action spaces, a harnessing agent guided by the optimized harness corrects the student's responses in the deployment action space before they run, and those corrected runs become the training demonstrations.

With the specialized harness removed, macro task success goes from 23.3% to 44.3%, higher than the 41.7% the base model reaches with the harness attached. Across 28 harness-induced behaviors in knowledge work, tool use, and science, 82.3% are recovered on average. For frontier models using the same evolved harness, the agent-as-harness form also beats the code-as-harness form. The paper is arXiv:2609.24974.

#harness-zero#agent-harness#google#ai-research

Best Taste-Bench model chooses the better direction 59.7% of the time

A paper from Microsoft and colleagues reports that the best model answers 59.7% of Taste-Bench questions correctly when choosing the better open direction in long engineering and research tasks. Forks decided by later evidence are much harder, a larger reasoning budget does not raise accuracy, and distilling a teacher’s outcome judgment improves held-out SWE-bench Pro success.