Intent alignment

Intent alignment asks one question: do an agent's tool calls stay inside the tools it was declared for? The score (0 to 100) compares declared tools to the tool calls the gateway saw, treating refunds.issue and refunds_issue as one tool. Model requests and MCP setup messages are not scored, and a session with no tool work has no score. The same number appears on the agent overview, the session page, and in HIRO. We never claim access to model-internal reasoning.

In practice: tag agents and assets, declare allowed tools in policy, then review mismatches in Work.

Intent alignment is not a guarantee that the model “thought” the right thing. It is an observable mismatch signal between what you declared and what the gateway saw.