Microsoft and Hugging Face release ThinkingBox, a benchmark grading AI agents on backend state consistency

Microsoft and Hugging Face released ThinkingBox, a benchmark that grades AI agents on the records and side effects they leave in backend systems rather than on their generated responses or tool calls alone.

The benchmark runs 507 stateful business workflows, each executed 20 times against various LLM-based agents, and reports three metrics: pass@1 (single-attempt success), pass@20 (at least one success in 20 tries), and observed 20/20 (tasks solved in all 20 attempts).

Across a common-set ablation covering 121,680 valid trials across 12 models, 79,853 attempts failed executable checks that inspect terminal backend state and side effects. Of those failures, 67.24% still terminated cleanly, invoked a state-changing tool, and reported no final tool error, indicating that surface-level correctness does not guarantee backend correctness.

Executable checks found wrong field values in 77.61% of failed attempts, unintended extra effects in 43.30%, and missing required effects in 25.36%, with these categories overlapping. The authors emphasize that a trajectory is a claim and database state is the evidence, with repetition serving as the trust test.

In single-attempt results (pass@1), Claude Opus 5.5 leads overall at 67.16%, followed closely by Claude Opus 5 at 66.50% and GPT-5.4 at 65.36%. Among open-weight models, Kimi-K3 leads at 57.37%. Domain performance varies widely: for example, Claude Opus 4.6 scores 68.62% on retail tasks but only 8.30% on auto insurance.

When evaluating repeatability across 20 trials, only three models retain most of their single-attempt score: GPT-6 Astra retains 78% of its pass@1 rate, while Claude Opus 5.5 and Claude Opus 5 each retain 71%. At the other extreme, GLM-5.1, Kimi-K2.6, and DeepSeek-V4-Pro each retain about 8%, illustrating a stark consistency gap.

Kimi-K3 demonstrates the broadest coverage, solving 93.89% of tasks at least once (476 of 507), but only 13.41% of tasks 20/20, underscoring that high breadth does not imply high consistency. The authors provide the benchmark and an example task for local execution via OpenEnv, enabling teams to reproduce and extend the evaluation.