Browse documentation

Set policies and runtime controls

Governance works best when policy reflects real business impact. Start with a small number of understandable rules, test them with representative actions, and expand only after reviewers can explain the outcomes.

Start with a default posture

Use this as a starting point, then adapt it to your systems and obligations:

Action patternSuggested outcome
Read-only, low sensitivity, known agent and systemAllow with recording
Internal update with limited, reversible impactAllow or review based on scope
Customer data, credentials, external communication, or production changeRequire review
Unknown identity, out-of-scope delegation, prohibited data, or unsafe targetBlock
Missing context needed to evaluate riskReview or block; do not silently allow

Give policy useful context

Consistent decisions depend on consistent inputs. Integrations should send stable values for:

  • agent and caller identity;
  • operation class and intent;
  • risk level and data sensitivity;
  • systems and resources touched;
  • environment;
  • delegation scope; and
  • relevant ownership or policy metadata.

Do not put secrets or unnecessary payload contents into policy metadata.

Design human review

Require a reviewer when judgment is more valuable than automatic speed. A useful approval request lets the reviewer see the actor, action, target, risk, reason, expiry, and recovery plan.

Set an explicit timeout. Your integration should stop safely when a request expires, is rejected, or cannot reach the approval service.

Protect data

Use data controls to block, mask, or review supported sensitive entities. Test with sanitized examples that represent the shape of production requests.

The test result is only as good as the context your integrations send. Confirm that each production path supplies the fields the policy expects.

Contain loops and spend

Rate limits, spend policies, and circuit breakers reduce the impact of retry storms, runaway agents, and unexpected usage. Set thresholds below the point where cost or downstream load becomes damaging.

Plan how the team will distinguish a legitimate traffic spike from a broken loop, and define who may restore service after a breaker opens.

Use layered controls

Roll out a policy safely

  1. Define the intended allow, review, and block cases in plain language.
  2. Test sanitized examples for each case.
  3. Start with one low-risk integration and named reviewers.
  4. Confirm reason details are understandable to operators.
  5. Exercise rejection, expiry, authentication failure, and service-unavailable paths.
  6. Monitor the Command Center after rollout.
  7. Expand by operation class or system, not by an unrestricted global exception.

Next step

Use the Command Center to operate the policy, review authored access in the Trust Graph, and connect Flight Recorder when your team needs deeper runtime investigation.