Skip to main content

Policy outcomes

Every action resolves into one of three outcomes:

Allow

Action is in scope and policy-safe.

Block

Action violates hard constraints.

Escalate

Action requires explicit operator approval.

Policy design guidelines

  • Use clear ownership boundaries per agent and environment.
  • Separate read-safe operations from state-changing operations.
  • Require approval for cross-environment writes and privileged tool calls.
  • Prefer explicit exceptions over broad allowlists.

Approval workflow basics

For escalated actions, reviewers should see:
  • Requesting agent identity and owner.
  • Requested tool action and inferred intent.
  • Relevant policy clauses and risk rationale.
  • Proposed blast radius and downstream impact.
Good approvals are evidence-first: reviewers should never approve without clear scope comparison and policy context.

Operating model

Start with conservative defaults and tune from trace feedback. Good policy systems evolve from observed behavior rather than fixed one-time rules.