Microsoft and Hugging Face have released ThinkingBox, an evaluation framework that grades AI agents on the records they leave behind — not the sentences they generate. The tool runs an agent against isolated tool sessions, then grades the terminal backend state and side effects it produces.
The core insight is straightforward: an agent that says "I booked the flight" is not the same as an agent that actually booked the flight. Most evaluation frameworks test whether the model's output looks correct. ThinkingBox tests whether the world changed the way the agent claimed it did.
"A tool call is not an outcome," the team writes. "One success is not reliability. Can you depend on the model behind your agent? What does consistency cost?"
ThinkingBox runs each agent through multiple episodes — the paper suggests twenty — and measures not just whether the agent succeeded, but whether it succeeded consistently. The framework produces a Pareto cost frontier: the tradeoff between price and reliability. An agent that works 95% of the time but costs twice as much as one that works 80% of the time is a different product, and buyers need to see that curve.
The framework is available through Hugging Face and can be run locally. It supports MCP tool sessions and OpenEnv servers, and the team has published the full technical report alongside the release.
For enterprises deploying AI agents, ThinkingBox addresses a gap that has been nagging at production teams: the difference between a demo that works and a system that works every time. If your agent books flights correctly 19 out of 20 times, that's not a 95% success rate — that's a customer service ticket waiting to happen.