🔍 Read the full analysis: The AI Agent Reported Success, But The Database Told Another Story on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Microsoft and Hugging Face have made ThinkingBox available through Hugging Face, offering a benchmark that checks the database state and side effects left by AI agents. Its authors report that many trials failed executable checks despite agents making state-changing tool calls without a final tool error; the results describe the benchmark setup, not performance across all business systems.
Microsoft and Hugging Face have made ThinkingBox available through Hugging Face, a benchmark that tests whether AI agents leave business systems in the required state after completing workflows. The release says it covers 507 tasks across five business areas and repeats each task 20 times, directing evaluation beyond plausible answers or valid tool calls to the recorded outcomes and side effects.
ThinkingBox runs agents in isolated sessions using Model Context Protocol (MCP) tools, then checks the resulting backend records against executable requirements. The benchmark covers retail, auto insurance, travel, neobanking and consulting. Its repeated runs are intended to measure consistency as well as whether an agent can succeed on a task at least once.
The authors’ example involves a delayed $745 appliance order. The customer does not qualify for late-delivery compensation under the policy the agent checks, but a carrier exception remains open. The required outcome is to keep the support ticket on hold while the issue is unresolved. Instead, the agent marks it solved and sends a reply that does not answer the customer’s underlying question. The executable check fails because the ticket is solved rather than on hold.
In a common-set analysis of 121,680 valid trials across 12 models, the authors report 79,853 attempts that failed executable checks. Of those failures, 67.24% ended without a final tool error even though the agent had invoked a state-changing tool. Checks identified wrong field values in 77.61% of failures, unintended extra effects in 43.30%, and missing required effects in 25.36%. These categories overlap, so they should not be added as separate shares of all failures.
Why Backend Outcomes Matter
For organizations assigning agents tasks such as refunds, support cases, insurance claims or bookings, a smooth conversation does not prove the requested work was done correctly. A system can accept a tool call while the agent records the wrong value, closes a case prematurely or makes an unintended change. Checking the final state can reveal those failures even when the interaction appears orderly.
The benchmark also makes a distinction between succeeding once and succeeding consistently. The release defines pass@1 as the share of attempts that pass, pass@20 as whether a task passed at least once in 20 runs, and observed 20/20 as passing every recorded run. These measures answer different questions: a single success among 20 attempts is not the same evidence as 20 successes. Neither measure, by itself, establishes reliability in a live company system.
In the authors’ reported table, Claude Opus 5.5 had an overall pass@1 of 67.16%, while Kimi-K3, identified as the strongest open-weight model in the table, scored 57.37%. The authors say Kimi-K3 was within one percentage point of GPT-6 Astra. Those figures are benchmark results, not a general ranking for every deployment or workflow.
Products Worth Considering
Bose S1 Pro+ Portable PA Speaker
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI-ARC, Black
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
As an affiliate, we earn on qualifying purchases.
From Tool Calls to System State
The release frames ThinkingBox around the gap between an agent’s actions and the state those actions leave behind. Valid tool use and a convincing final response are indirect signals: they do not necessarily show that a task’s requirements were met. The benchmark instead evaluates backend records and side effects after each run, using isolated sessions and a clean starting state.
The authors describe the release as based on their paper and say the benchmark can be run through OpenEnv. The material provided does not give a publication date or the full evaluation details, including all task specifications and model configurations. It also does not report independent replications. The findings should therefore be read as the authors’ results under their tested setup, rather than as an estimate of how agents perform across all business systems.
“A tool call is not an outcome.”
— Microsoft and Hugging Face, in the ThinkingBox release
Limits of the Benchmark Results
The supplied release material does not specify when ThinkingBox became available or provide the full paper details needed to assess the evaluation in depth. It gives no uncertainty estimates for the reported model scores, so the stated differences should not be read as proof that one model is reliably better than another across settings.
It is also unclear whether the 507 workflows represent the range of tasks, policy exceptions and changing records found in any particular organization. The authors’ results do not establish how the models would perform in live systems or under different configurations. Passing all 20 observed runs is evidence limited to those trials; the supplied material does not show that it predicts long-term reliability or performance on workflows outside the benchmark.
How Teams Can Test Their Workflows
The release says researchers and developers can run ThinkingBox through OpenEnv using isolated MCP tool sessions and compare outcomes against executable checks. It does not identify a planned follow-up date, independent replication or other upcoming milestone.
For businesses considering agents, a practical next step is to test workflows that match their own policies and systems, then inspect both passing and failing runs. That can help identify whether an agent records the required values, avoids unwanted side effects and handles unresolved cases appropriately. Whether results on ThinkingBox predict performance in a particular organization remains to be tested.
Key Questions
What does ThinkingBox evaluate?
ThinkingBox checks whether an AI agent leaves backend records and side effects in the state required by a workflow, rather than judging only its tool calls or final response.
How large is the benchmark?
The release describes 507 business workflows across retail, auto insurance, travel, neobanking and consulting. Each task is repeated 20 times from a clean backend.
What did the authors report about failures?
In a common-set analysis covering 121,680 valid trials across 12 models, 79,853 attempts failed executable checks. The authors say 67.24% of those failures ended without a final tool error despite a state-changing tool call.
Do the scores predict how an agent will perform at my company?
Not on their own. The results describe the authors’ tested benchmark setup. Performance in a company’s own systems remains uncertain and depends on its workflows, data, policies and operating conditions.
Can developers run ThinkingBox themselves?
The release says it can be run through OpenEnv using isolated MCP tool sessions. The supplied material does not give a future release date or name an independent replication.
Primary source: Hugging Face · via ThorstenMeyerAI.com
