By Hamza Faisal
Microsoft and Hugging Face have teamed up to answer an uncomfortable question in the age of AI agents: did the agent actually do the work, or did it just say it did?
The result is ThinkingBox, a new evaluation framework introduced in a joint blog post on October 3, 2026, that grades agents not on what they say in their final message, but on what they wrote into the database behind them.
The motivating example is striking. A support agent handling a late delivery makes nine well-formed tool calls, reads the refund policy correctly, and closes the ticket as "resolved."
A grader looking only at the tool calls or the final sentence would mark it a success. But the courier exception was still open, meaning the required end state was "on hold," and the customer never got an answer to her actual question.
The executable check that fails is a single database field.

How the benchmark works

ThinkingBox-Bench runs 507 stateful business workflows across retail, auto insurance, travel, neobank, and consulting domains, each repeated 20 times per model against 18 proprietary and open-weight large language models.

Every attempt starts from an identical clean backend. Of the 507 tasks, 477 are graded on database state alone, with 30 adding response rubrics.
In a common-set analysis of 121,680 valid trials across 12 models, 79,853 attempts failed the executable checks — even though 67.24% of those failures terminated cleanly with no reported tool error.
The failure breakdown is revealing: 77.61% of the failures had wrong field values, 43.30% produced unintended extra effects, and 25.36% missed a required effect entirely.

One success is not the same as dependability


The results show that single-attempt accuracy and real-world reliability are very different things. Claude Opus 5.5 leads overall first-attempt accuracy at 67.16%, with Kimi-K3 the strongest open-weight model at 57.37%.

But consistency tells another story: Kimi-K3 solves 93.89% of tasks at least once yet passes only 13.41% — 68 of 507 — on every single attempt, while Claude Opus 5 solves fewer tasks at least once (79.09%) but passes 47.53%, or 241 of 507, every time.

Even more striking, the newer Claude Opus 5.5 scored higher on average than Claude Opus 5 but passed exactly the same number of tasks — 241 — on all 20 attempts, meaning its accuracy gain bought no added dependability.

The authors estimate the cost of the full campaign at $7.80 per dependable task for Opus 5.5 versus $13.30 for Opus 5, at list rates.

Why it matters for real deployments

The implication for teams building agentic systems is hard to miss: evaluations that look great in testing can hide problems in production. An agent that quietly leaves wrong values in a database is far more dangerous than one that errors loudly, because downstream systems treat those wrong values as ground truth.

By checking backend state and side effects rather than the agent's own confident reply, ThinkingBox aims to push the industry toward evaluations that match what production actually feels like — where the ticket really is on hold, the refund really went through, and the customer really got her answer.

By Hamza Faisal