Microsoft’s ThinkingBox turns database checks into rewards for training agents
Microsoft and Hugging Face described ThinkingBox on 22 October, a sandbox and benchmark built by Microsoft's Copilot Studio team with Toloka that grades agents by the final state of a real database rather than by their conversation. ThinkingBox-Bench has 507 tasks, each with a real backend, MCP tools and a simulated user; deterministic checks on the end state return 1 or 0, which the team uses directly as a reinforcement learning reward.
They trained Qwen3.8-27B with GRPO for 50 steps on 187 tasks of middling difficulty, chosen because tasks that are always or never solved give no learning signal. Pass@1 rose from 51.70% to 60.78%. The training code is on GitHub, ThinkingBox is MIT-licensed, and the benchmark data is under CDLA-Permissive-2.0; the public tasks are synthetic.
Checking the state an agent leaves behind is a more honest test than reading its transcript, and the same check doubles as a training reward. Start with the three actions your agent takes most often.
Drafted with AI from the original source, then checked and edited by Sridhar Mukkandi. Spotted a mistake? Write to hello@tarkika.com and we'll correct it here.