An AI agent can finish a task, report success and still leave the database in the wrong state. Microsoft’s Copilot Studio team built a benchmark around that gap.
ThinkingBox runs agents against isolated tool sessions, then grades the records and side effects they leave behind instead of the words they produce. Its library holds 507 stateful business workflows. Every task runs 20 times from a clean backend, so a single lucky pass cannot hide deeper unreliability.
The scale of the problem shows in the numbers. Across 121,680 valid trials spanning 12 models, 79,853 attempts failed the checks. Roughly two-thirds of those failures ended cleanly with no tool error, yet the checks caught wrong field values in 77 percent, extra effects in 43 percent and missing effects in 25 percent.
Consistency separates the field. Claude Opus 5.5 led at 67.16 percent on a single attempt. Only three models kept most of that score across 20 repeats. Kimi-K3 solved the most tasks at least once but lasted all 20 tries on just 13 percent of them.
The authors say most failures come down to tool handling and recovery, not reasoning. The harness and benchmark are now available on Hugging Face.
