Microsoft Open Sources Thinkingbox: Stateful Agent Workflow Benchmark
microsoft · hf · 2026-08-25
Microsoft released Thinkingbox, a sandbox and benchmark for evaluating agents in stateful business workflows. Unlike single-turn benchmarks, it requires multi-turn information gathering, policy adherence, and correct persistent state transitions.
- Scenarios: Covers 507 policy-conditioned workflows across retail, hospitality, auto insurance, neobank IT, and consulting.
- Evaluation: Uses task-specific executable checks to validate trajectories and reject side effects.
- Findings: The top model achieves 65.36% pass@1 but only 25.25% pass@20. Many failures involve valid-looking actions, revealing that single-step signals are poor proxies for end-to-end reliability.
Code and benchmarks are open-sourced on GitHub.
More from coding & agent
- Tool to support BYOA and host custom agents — tekbog · 2026-08-25
- New Agent Platform Offers Full Computer Control and Custom Agents — basedjensen · 2026-08-25
- Discussing observability stacks for LangChain agents in production — Human-Agent6509 · 2026-08-25
- RAW Sequencer: A Desktop Tool to Compare Up to 6 AI Video Generations Side by Side — AniketBhatkar · 2026-08-25
- Multi-agent architecture: integrated workspace vs thin coordination layer? — GlitteringAngle8601 · 2026-08-25
- How to create reusable AI skills by reverse-engineering examples — skillfusion_ai · 2026-08-25