Microsoft tested AI agents across 507 business workflows and found most failed behind the scenes

Microsoft released ThinkingBox on Hugging Face, an evaluation framework that judges autonomous AI agents by the records they modify rather than the sentences they draft.

Most benchmarks reward an assistant if it issues valid tool queries and writes a helpful reply. ThinkingBox takes a different route. It inspects whether the database actually reflects the intended change after the model finishes running.

The findings show that polite language often masks broken work. Across an evaluation of 12 frontier models over 507 business tasks, the researchers analyzed nearly 80,000 failed attempts. In over 67% of those failures, the agent reported no technical error and ended its conversation normally. Yet deep database checks revealed wrong values in over 77% of those cases, alongside unexpected extra changes in 43%.

Performing a task correctly once also proved very different from repeating it reliably. Each scenario was tested 20 independent times from a clean state. The open-weights model Kimi-K3 demonstrated broad reach by solving almost 94% of tasks at least once, but it managed to pass all 20 repeats on just 13.41% of them. Meanwhile, Claude Opus 5 solved fewer unique tasks but completed 47.53% of the catalog without a single slip across 20 tries.

Consistency also reshapes the economics of deployment. GPT-5.6 Sol delivered the lowest cost for an isolated successful run at $0.127, but achieving a completely dependable 20-run task raised that expense to $9.76. In comparison, GPT-5.4 reached dependable status on 128 tasks at $6.80 each, while GPT-6-Astra reached 231 consistent tasks at $7.45 each.

The data shows that roughly 80% of all agent breakdowns stem from poor tool handling and an inability to recover from minor technical errors, rather than a failure of logic. An agent that meets an empty database lookup often gives up and presents the task as solved with a cheerful smile.

Would you let an automated agent manage financial refunds if it quietly wrote the wrong status to your database one time out of five?

The benchmark currently focuses on simulated business operations, with tasks in auto insurance scoring far lower on average (around 34%) than retail workflows (around 60%). Microsoft also noted that it has not yet measured how much human oversight or smart error-retry logic might raise these scores in real production software.