In a joint blog post published October 3, 2026, Microsoft and Hugging Face introduced ThinkingBox, a benchmark that grades AI agents on the records they leave behind in a backend database rather than on the tool calls they make or the final messages they write. The post, authored by Microsoft's Tuhin Kundu with contributors from Enderis AI, Hugging Face and several university interns, is now available through Hugging Face via OpenEnv. ThinkingBox covers 507 stateful business workflows, each run 20 times against a clean backend, grading agents on terminal backend state and side effects. According to the post, in a common-set ablation covering 121,680 valid trials across 12 LLM models, 79,853 attempts failed executable checks. Of those failures, 67.24% still terminated cleanly, invoked a state-changing tool, and reported no final tool error. Executable checks found wrong field values in 77.61% of those cases, unintended extra effects in 43.30%, and missing required effects in 25.36%. On pass@1, Claude Opus 5.5 led at 67.16%. When tasks were repeated 20 times, GPT-6 Astra retained 78% of its single-attempt rate, while GLM-5.1, Kimi-K2.6 and DeepSeek-V4-Pro each kept about 8%. The authors note domain matters: Claude Opus 4.6 scored 68.62% on retail but 8.30% on auto insurance.
- 507 stateful workflows, each run 20 times against a clean backend, grading final database state
- Of 79,853 failed trials, 67.24% still terminated cleanly and reported no error
- Consistency varies widely: GPT-6 Astra kept 78% of its single-attempt score; some models kept ~8%
What it means for you
An AI agent can confidently say a task is done while the underlying system shows it wasn't—wrong value saved, ticket closed instead of held, an extra action taken. This benchmark measures that gap by checking what the agent actually changed, and by running each task 20 times to see if it works reliably or just got lucky once. The headline lesson: a demo that works once is not proof an agent works every time.
Try this
If you're testing or running an AI agent that touches real systems (orders, tickets, records), check its outcomes against the actual data it changed—not its own 'done' message—and run the same task several times to see if results stay consistent.
Who should care
Anyone building, buying, or evaluating AI agents that write to real business systems like CRMs, ticketing tools, or order databases.
Skip this if
You only use AI for drafting, summarizing, or chat, or you don't yet have agents making changes to live systems.
Sources: Hugging Face — read the original