Morning Edition · Thursday, October 8, 2026Published at 2:25 AM EDT · New York

Google and Cambridge's VeriHarness uses the generating model to audit its own agent rollouts and adds about 6 points on long-horizon tasks

The method treats agreement between rollouts as suspect and disagreement as a clue, and the authors have released the code with about 26,000 agent traces.

Researchers at the University of Cambridge and Google Cloud AI Research have open-sourced VeriHarness, a harness for checking the results of long-running agent tasks. It needs no reference answers or grading rubrics at test time. According…