Eigen RadarAI
Archive
Analysis

ThinkingBox tests the business-system results agents leave behind

Microsoft and Hugging Face made ThinkingBox available, with 507 workflows that check what AI agents actually change in business systems after saying a task is finished. Isolated tool sessions let executable checks identify missing actions and unwanted extra changes. Researchers tested 18 models over 20 repetitions of each task, distinguishing occasional task completion from consistently getting the workflow right.

Artificial Intelligence··Morning
A monitor seen from behind, connected desktop computer, keyboard and mouse on a wooden office desk.

ThinkingBox inspects the state left after an agent finishes

Microsoft and Hugging Face made ThinkingBox available through Hugging Face, a platform hosting machine-learning tools and datasets. The evaluation environment checks AI agents against the business-system state they leave after using software tools. Its 507 workflows cover retail, auto insurance, travel, digital banking and consulting. An agent’s declaration that a job is finished is checked against the actual changes in those systems.[1], [2]

Each repetition begins in an isolated tool session with a clean starting state. A simulated user holds private context and provides some of it only when the agent asks an appropriate question. Grading checks required field values, missing updates and unintended extra actions. Of the tasks, 477 are graded solely on state changes; another 30 also include criteria for the final response. The domain counts are 98 retail tasks, 100 insurance tasks, 104 travel tasks, 104 digital banking tasks and 101 consulting tasks.[1]

A delivery complaint is closed when it should remain on hold

The developers illustrate the test with a simulated customer whose kitchen appliance is held at a Nashville distribution center, 15 days late. Nashville is a US city. The agent makes 9 tool calls and marks the complaint resolved. The task rules require keeping it on hold. The error is in the resulting complaint status, despite the agent having performed a sequence of tool actions.[1]

Repeated trials separate task completion from consistency

The research team tested 18 models, repeating each task 20 times. A common-task analysis of 12 models contains 121,680 trials, with 79,853 classified as unsuccessful. The team reports that 67.24 percent of unsuccessful trials ended cleanly, changed state and encountered no tool error. Wrong values, extra side effects and missing changes can overlap as failure categories, while operational failures remain unsuccessful outcomes.[1]

The environment connects to OpenEnv, Hugging Face’s infrastructure for interactive evaluation environments. It supports evaluation, training and experiments with different tasks, with framework and dataset versions pinned. The developers measure both correct workflow completion and consistency across repeated attempts. Their published numerical results describe experiments in this environment, rather than observations of real customers completing work with deployed systems.[1]

References

  1. News sourceMicrosoft / Hugging FaceThinkingBox opens tests of the results AI workflows leave behind↩1↩2↩3↩4↩5
  2. News sourceSuperpower DailyMicrosoft opens ThinkingBox evaluation through Hugging Face↩
AI safety and regulationChip wars and the supply chain

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →