ThinkingBox is an open-source test bed from Microsoft that grades AI agents on what they actually change in a system, not on what they say they did.
In plain words
An agent can make tidy tool calls and still leave a database wrong. ThinkingBox connects the agent to isolated tool servers built on MCP, then checks the final state of the backend and any side effects. Its benchmark has 507 business workflows, and each one is run 20 times per model, so it measures whether an agent succeeds every time rather than once. The code is MIT-licensed, and it can also be used to make training data and train agents with reinforcement learning.
A quick example
Imagine testing a customer-service agent that should issue a refund. It replies "Done!", but the test opens the order records and finds no refund was saved, so that run counts as a failure.
Where you'll meet it
- Measuring whether an AI agent is reliable enough to use on real business tasks.
- Comparing models by how often they succeed on every attempt.
Related
Sources
- DocsThe Agent Said It Was Done. The Database Disagreed., Microsoft (Hugging Face blog) · read 4 Oct 2026
- Docsmicrosoft/thinkingbox, Microsoft (GitHub) · read 4 Oct 2026
Checked by WhatIsAI against the sources above · 4 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.