Mouse-1 Turns Agent Choices Into API

Mouse-1 targets the overlooked control-flow layer of agents, where cheap, fast decisions can reduce frontier-model latency and cost without replacing deeper reasoning.

• –Scores tool routing, action gating, verification, and human-review decisions directly
• –Reports 89.1% accuracy on unfamiliar tools and repositories, though it trails d3 and Clef on its broader tool-and-security suite
• –Its claimed 0% decision-flip rate under planted prompt injection is promising, but not a substitute for end-to-end security testing
• –The strongest use case is pairing Mouse-1 with a frontier model: Mouse-1 handles repetitive branching while the larger model handles planning and code
• –Developers should validate calibration and failure costs inside real agent loops before trusting probabilities for irreversible actions