The Agent Said It Was Done. The Database Disagreed.
Hugging Face Blog 发布了这条动态，详情请查看官方发布。
正文 · 来源原文
Figure 1: ThinkingBox runs an agent against isolated MCP tool sessions, then grades the terminal backend state and side effects it leaves behind. From our ThinkingBox paper.
This is a joint blog by Microsoft and Hugging Face, special thanks to Tommy Guy (founder at Enderis AI, previously Microsoft), Sergio Paniego from Hugging Face and our former interns Zhuochun Li (Univ
A customer writes in. Her $745 kitchen appliance has been stuck in a courier "exception" at a Nashville distribution center, fifteen days past its estimated delivery date.
The AI agent does careful work. Nine tool calls: it pulls the order, checks tracking, looks up her customer profile, searches the refund policy twice, confirms no ticket exists, opens one, documents the timeline, and reads the policy correctly; her account segment genuinely does not qualify for late-delivery compensation.
Then it closes the ticket as resolved and replies “ Since your query is resolved, is there anything I may assist you with? ”
Two things are wrong. The carrier exception is still open, so the required end state was on hold, pending resolution. And the customer never got a real answer to what she actually asked.
Claude Opus 5.5 总体领先，达到 67.16%，比 Claude Opus 5 高出三分之二。Kimi-K3 是最强的开放权重模型，与 GPT-6-Astra 相差一个百分点。领域同样重要：Claude Opus 4.6 在零售方面得分为 68.62%，在汽车保险方面得分为 8.30%。
这三者中没有一个占主导地位：每增加一项可靠的任务都会花费更多。 Claude Opus 5 也超过了 241，但价格为 13.30 美元，因此 Opus 5.5 完全占据主导地位。 GPT-5.6 Sol 的单次成功成本最低，为 0.127 美元，而每个可靠任务的成本为 9.76 美元。获得正确答案的最便宜方法并不是获得可靠答案的最便宜方法。
Classify tool and system errors so retries target the recoverable ones.根据工作流程需要切割工具表面。 And require human approval on the changes you cannot cheaply reverse. We have not measured the lift from any of these on this benchmark, which is exactly the kind of thing the environment now makes testable.
ThinkingBox 现在已在 Hugging Face 上，包括线束和数据集。 ThinkingBox-Bench 现在位于 OpenEnv 界面后面，每个完成的剧集都会返回二进制通过/失败奖励。发布的适配器专为评估而设计；单独的非基准场景可以在训练工作流程中使用相同的界面。
在 Linux 和 WSL 上使用 Python 3.11+、uv 和 Docker 进行了测试。您还需要在固定的发布和模型端点处为代理、模拟用户和法官进行思考盒数据检查。一个端点可以服务所有三种角色，这是最简单的开始方式。 The OpenEnv image starts only the OpenEnv API; everything else you run yourself.
An AI grader checking tool calls would see nine well-formed ones. The grader checking whether the agent wrote to the database would see that too. The database is what disagrees.
That gap is what ThinkingBox measures. Across 507 stateful business workflows, each run 20 times against various LLM models, it grades agents on terminal backend state and side effects. This post covers what we found, what consistency costs, and how to run the benchmark yourself through OpenEnv.
You can run this one yourself: the example above is adapted from a benchmark task sandbox_external_retail_group1.py:test_case_ST003_006, and the executable check that fails is a single field: the ticket's status is solved where the required end state is hold. The full trace is in Appendix D.4, Case 3 of our paper.
Contents
A tool call is not an outcome One success is not reliability Can you depend on the model behind your agent? What consistency costs Failure signatures How it works Run it yourself Where this goes next
Want to try it before reading the results? Skip to section Run it yourself.
Final responses and valid tool calls are only proxies. An agent can sound correct while leaving the wrong value, changing the wrong record, or creating an extra side effect. Only the records it leaves behind settle the question.
The gap is substantial. In a common-set ablation covering 121,680 valid trials across 12 LLM models, 79,853 attempts failed the executable checks. Of those failures, 67.24% still terminated cleanly, invoked a state-changing tool, and reported no final tool error. Executable checks nevertheless found wrong field values in 77.61% of them, unintended extra effects in 43.30%, and missing required effects in 25.36%. Those state-check findings overlap.
A trajectory is a claim. Database state is the evidence. Repetition is the trust test.
An agent that processes a refund correctly once and mishandles it the next four times is not a working refund agent. So every task runs 20 independent times, each from an identical clean backend, and we report three different things:
Table 1: The three numbers we report, and the question each one answers.
We use observed 20/20 in this blog post as the literal count of how many of the 507 tasks passed 20 out of 20. No estimator, no smoothing.
Starting with the familiar view. The table below reports pass@1, the single-attempt score estimate, broken out by domain. This is the number most leaderboards publish, and on its own it reads like an ordinary capability ranking.
Table 2: ThinkingBox-Bench pass@1 (%) by domain. Each model is evaluated on every task for 20 repeated trials. Bold marks the group leader; underline marks the runner-up. The standard errors for the single attempt score estimates are provided in Table 4 in our ThinkingBox paper.
Claude Opus 5.5 leads overall at 67.16%, two-thirds of a point above Claude Opus 5. Kimi-K3 is the strongest open-weights model, within a point of GPT-6-Astra. Domain matters just as much: Claude Opus 4.6 scores 68.62% on retail but 8.30% on auto insurance.
One good run tells you a model can do the work. It does not tell you whether it will do it again. So run every task 20 times and ask how much of that score survives.
Figure 2: How much of each model's single-attempt score survives 20 repeats.
Only three hold on to most of their pass@1 scores: GPT-6 Astra retains 78% of its single-attempt rate, and Claude Opus 5.5 and Claude Opus 5 each retain 71%. At the other end, GLM-5.1, Kimi-K2.6 and DeepSeek-V4-Pro each keep about 8%.
The gap between what a model can do once and what it does every time is the whole story.
Figure 3: Breadth and consistency pull apart. Twelve of the eighteen models are shown; six below 33% pass@1 are omitted for legibility.
Kimi-K3 has the broadest coverage of any model we tested. It solves 93.89% of the benchmark at least once: 476 of 507 tasks. Only 31 tasks defeat it entirely, the lowest count in the field. On retail workflows it leads outright at 82.24% pass@1, ahead of every proprietary model.
Kimi-K3 is also among the least consistent. Just 68 of 507 tasks, 13.41%, succeed in all 20 attempts.
Claude Opus 5 inverts this. It solves fewer tasks at least once (79.09%; 106 defeat it entirely) but completes 47.53% of the benchmark on every single attempt.
A newer model does not fix this. Claude Opus 5.5 scores higher than Claude Opus 5 on every-attempt average, 67.16% against 66.50%, and solves more tasks at least once. It passes exactly the same number of tasks on all 20 attempts: 241. Half a point of headline accuracy bought no additional dependability at all.
Kimi-K3 solves 75 more tasks at least once than Opus 5. Opus 5 solves 173 more tasks consistently than Kimi-K3.
If you are choosing a model for work that touches real records, pass@20 is the wrong column to look at.
Capability comparisons usually stop at the score. For anyone deploying, the relevant question is what a successful unit of work costs. We measure that as cost per successful task attempt. We say task attempt because every benchmark task is run repeatedly and cost is incurred per attempt, so pass@1 is the matching quality denominator.
We took each model's recorded token usage from its full 507 × 20 campaign and priced it at undiscounted list rates available on OpenRouter+, reversing promotional discounts and excluding endpoints that declare quantization. Input, output and cache rates all come from one provider endpoint per model.
Then we divided one run's cost by the number of attempts that succeeded:
Cost per successful task attempt = estimated cost for 507 attempts, one per task ÷ (507 × pass@1)
This is a comparative efficiency index, not an invoice, and not the price of serving one production request. It also prices single successes, not consistency. We price consistency next.
Example: GPT-5.4 costs $43.49 for 507 attempts (one attempt per task) and has 65.36% pass@1, so $43.49 ÷ (507 × 0.6536) = $0.131 per successful task attempt.
A model is on the frontier if no other model is both no more expensive and at least as accurate. Three models qualify; every other model is dominated on at least one axis.
Figure 4: Cost per successful task attempt against pass@1. Ringed dots are pareto cost frontier models.
The frontier has three steps. GPT-5.6 Sol has the lowest cost per success at $0.127; GPT-5.4 raises pass@1 by 3.45 percentage points for $0.004 more per success; Claude Opus 5.5 adds another 1.80 points at $0.276 per success. Each of the three remains on the cost frontier line because no cheaper model matches its pass@1.
Claude Opus 5 is the clearest case: at $0.475 per successful attempt and 66.50% pass@1, it is both more expensive and less accurate than Claude Opus 5.5 at $0.276 and 67.16%.
Cost per success rewards a model that is cheap and often right. It does not reward a model that is right every time. So we also compute cost per dependable task: the cost of the full 20-run campaign divided by the number of tasks the model passed on all 20 attempts.
Cost per dependable task = estimated cost of 20 runs of 507 attempts ÷ tasks passing 20/20
Example: GPT-6-Astra costs 20 × $86.03 = $1,720.60 for the campaign and passes 231 tasks on every attempt, so $1,720.60 ÷ 231 = $7.45 per dependable task.
Table 3: The nine lowest costs per dependable task among models with at least one observed 20/20 task, sorted low to high. Estimated $, not actual cloud bills.
Now rank by consistency. GPT-5.4 is the cheapest at $6.80, though only 128 tasks meet the bar. GPT-6 Astra reaches 231 at $7.45, and Claude Opus 5.5 the joint-highest 241 at $7.80.
None of the three dominates the others: each additional dependable task costs more. Claude Opus 5 also passes 241, but at $13.30, so Opus 5.5 dominates it outright. GPT-5.6 Sol, the cheapest per single success at $0.127, costs $9.76 per dependable task. The cheapest way to get a right answer is not the cheapest way to get a dependable one.
We assign each failed trace one deterministic diagnostic signature, and the headline is actionable: roughly four in five failures are tool handling, not reasoning. Across an ablation study in Table 5 of our paper:
These are unweighted averages of per-model shares and observable labels, not unique causal explanations.
The practical pattern is simple: agents usually get far enough to attempt the workflow, then fail to recover from tool errors, failed preconditions, or empty lookups. That is a retry and error-recovery problem before it is a model problem.
Difficulty also changes by domain: across the models listed in Table 2 above, retail averages 59.52% pass@1 while auto insurance averages 33.83%.
What to do about it. Treat the 20/20 rate as a design input, not a verdict. The same signal the benchmark grades on is available in production: check the terminal state before you commit, not the model's summary of it.
Classify tool and system errors so retries target the recoverable ones. Cut the tool surface to what the workflow needs. And require human approval on the changes you cannot cheaply reverse. We have not measured the lift from any of these on this benchmark, which is exactly the kind of thing the environment now makes testable.
ThinkingBox is the agent sandbox, while ThinkingBox-Bench is a dataset benchmark to evaluate agents. The diagram at the top of this post shows the loop; here is what each part does.
Figure 5: The sandbox loop from panel A in Figure 1 above: isolated tool session, terminal database state, side effects, executable judges.
Each task defines a starting backend state, a user goal, the available MCP tools, the domain policy, and executable checks over the terminal state. A simulated user holds private context (a booking reference, a preference, or a date of birth) and releases it only when asked.
Every attempt gets an isolated MCP session with freshly initialized state. Two attempts of the same task never share a database row or cached tool state, which is what makes 20-trial comparison meaningful.
At the end, a side-effect extractor derives what actually changed, and deterministic judges compare it against the required end state, accepting any trajectory that produces the right outcome while rejecting wrong, missing or extra effects. For requirements with no clean database value ("did the agent disclose this is not guaranteed?"), a narrow binary rubric question handles the semantics. 477 of the 507 tasks are graded on state alone; 30 add response rubrics.
The trust boundary: the model sees tasks, dialogue and tool schemas. Golden state, assertions, grading internals and credentials stay on the evaluator side.
ThinkingBox is now on Hugging Face, both the harness and the dataset. ThinkingBox-Bench now sits behind the OpenEnv interface, and each finished episode returns a binary pass/fail reward. The released adapter is designed for evaluation; separate, non-benchmark scenarios can use the same interface in training workflows.
Tested on Linux and WSL, with Python 3.11+, uv and Docker. You also need a thinkingbox-data checkout at the pinned release and model endpoints for the agent, simulated user and judge. One endpoint can serve all three roles, which is the simplest way to start. The OpenEnv image starts only the OpenEnv API; everything else you run yourself.
In a second terminal, start Typesense 30.1 and wait for its health check:
In a third terminal, start the Session Proxy and MCP servers.
Back in the first terminal, start the OpenEnv server against a ThinkingBox YAML config naming your three models (config guide):
Gate on readiness before running anything. It returns 503 until its observable data, configuration and Session Proxy checks pass. It cannot observe Typesense or live-probe every model endpoint, so confirm those separately:
Now score a real episode. example_usage.py only resets and lists tools; for agent actions, effects and assertions use the packaged evaluator:
The OpenEnv adapter writes operational failures to an errors sidecar so they can be rerun rather than silently mixed with model outcomes. A canonical result must resolve or explicitly account for those attempts; we counted system errors as unsuccessful trials.
Runs are gated on a pinned framework commit, a pinned data release and a bundle hash, so a canonical result is verifiable rather than asserted.
The useful part of this work is not our pass@1 leaderboard. It is the environment.
If you are evaluating an agent that touches real records:
Inspect a failure. Find a run that terminated cleanly and still failed, and look at what actually changed in the database. It reframes what your own evals measure. Reproduce one task through OpenEnv with your own model. Report a repeat metric, and define it. Whatever k your use case justifies; say whether you are reporting best-of-k or every-of-k, and how you computed it.
Further details can be found in the following links:
ThinkingBox code is MIT-licensed; the benchmark data is CDLA-Permissive-2.0; the OpenEnv environment ships under OpenEnv's BSD-3-Clause.
Disclaimer: every task in the public benchmark is a synthetic reconstruction. The workflows and policies are modeled on real AI agentic enterprise patterns; the customers are not real.
ThinkingBox and ThinkingBox-Bench is built by the Microsoft Copilot Studio team in partnership with Toloka with collaborators from the University of Pittsburgh, Northwestern University, Columbia University and UC Irvine who interned at Microsoft. Questions are welcome in the comments section below or on github
+ OpenRouter cost snapshot taken on Sept 20th 2026; Opus 5.5 pricing as per the Anthropic site.
