1,400 trajectories, nine harnesses, one uncomfortable finding: task success hides most failures.
The State of Agent Evaluation, 2026
Research · Project
We ran 1,400 agent trajectories across nine public harnesses and found that task success alone hides most failures. This report proposes a four-layer rubric.
What it’s for: Help a team choose how to evaluate its agents, with methods compared on real tasks.



