How to Measure If Your AI Agent Is Safe to Run Unsupervised Overnight

A recent benchmark tested 507 state-mutating workflows across 20 attempts each, revealing a stark gap between single-attempt accuracy and consistent reliability — the top model succeeded 66.5% of the time per attempt but achieved perfect consistency on only 47.53% of tasks. Developers are advised to evaluate agents using a metric called pass^k, which checks whether an agent completes a task correctly every single time, not just occasionally. The approach requires tasks that mutate stored state — such as processing refunds or updating records — since read-only tasks cannot expose the failure modes that matter. A proper test harness must include full environment resets between runs and assertions that verify both required changes and the absence of unintended side effects. The benchmark found that most tasks fall into a dangerous middle category where agents succeed only sometimes, making silent data corruption a real risk that standard observability tools like error rates or latency metrics would fail to detect.