Agent-Safe Pipeline binds agent actions to grants

Decionis published a reference pipeline in which an agent proposes an action, trusted code captures an expiring intent, an independent policy gate returns ALLOW, ESCALATE, or BLOCK, and a SafeExecutor consumes a single-use authorization before calling a registered handler. Why it matters: Human approval is a weak checkpoint if agent output can also choose the identity, endpoint, or credentials. Binding an authorization to the captured intent helps stop approval for one action from authorizing another.

Try this: Run the GitHub-deploy agent example with an expired or altered intent, then inspect the blocked trace and confirm that the executor never reaches its registered handler.