0% read on this device
Browse the curriculum

Start Here

Vision

First Principles

Operating Model

Domain Model

Agent Factory

Runtime Architecture

AI Engineering

Autonomous Workflows

Verification & Delivery

Factory Platform

Quality Engineering

Security & Governance

Case Studies

Labs

Interview Practice

Research Journal

Reference

Curriculum/Security & Governance/Complete source chapter
Security & Governance5 min readchapterQuick Read

Agentic Governance Control Framework

Convert governance principles into ten testable control families.

Status: Review readyRisk: criticalLifecycle: intent · plan · execute · verify · deliver · learnContent reviewed 2026-08-30Maturity guide →
Claim boundaryThis is curriculum guidance. It does not by itself prove a production implementation.
Quick Read

The chapter in one pass.

~2 min
  • Purpose: Convert governance principles into ten testable control families.
  • Control contract: Owner, intent, enforcement point, evidence, test, exception, failure response, and review cadence.
  • Core rule: A policy document or visible button is not proof that a control prevents, detects, contains, or recovers from failure.
  • Maturity: Review-ready control design; operating effectiveness must be established in the target system.

1. The problem

Autonomous systems connect probabilistic decisions to tools that change real systems. Risk compounds across delegated steps, external content, shared memory, suppliers, and retries. Governance that reviews only the final output cannot reconstruct authority, identify the failed control, or contain a live incident.

2. Enduring Principle

Govern every transfer of authority

Controls attach to identities, records, grants, state transitions, tool calls, artifacts, evidence, approvals, releases, incidents, and learning promotions. They record observable decision and action lineage, not hidden model reasoning.

3. Control catalog

IDControl familyOwnerPrimary enforcementRequired evidenceTest and cadence
GOV-01Authority-chain mappingGovernance ownerControl plane and policy serviceActor, delegator, grant, scope, expiry, decision chainTrace one high-risk action quarterly and after authority changes
GOV-02Unique identityIdentity ownerIdentity provider, workload issuer, tool gatewayHuman/service/workload/agent/capability identities and credential eventsDeny shared, expired, and wrong-scope credentials each release
GOV-03Risk and autonomy tiersSystem ownerAdmission and policy decision pointsClassification, ceiling, allowed/prohibited actions, approval ruleBoundary cases and promotion regression each policy change
GOV-04Emergency controlIncident authorityOrchestrator, scheduler, gateway, credential brokerPause, cancel, revoke, quarantine, rollback, recovery recordsFailure injection at risk-based cadence
GOV-05Agentic attack resistanceSecurity ownerIngestion, context compiler, model/tool gateway, sandboxSanitization, provenance, policy denials, alerts, containmentIndirect-instruction, tool, memory, context, and supply-chain attacks each release
GOV-06Human override and dual controlDecision ownerApproval service and protected operationNamed approvers, decision, reason, counterevidence, timingReject self-approval and missing second party quarterly
GOV-07Decision and delegation lineageAudit ownerEvent and evidence pipelineCorrelated inputs, versions, actions, outputs, decisions, delegationsReconstruct sampled runs and every material incident
GOV-08External capability diligenceCapability ownerIntake and registry admissionOwnership, provenance, data use, security, service, exit, incident termsIntake plus periodic recertification and material supplier change
GOV-09Incident reportingIncident ownerDetection and case managementClassification, timeline, scope, notification decisions, postmortemTabletop twice yearly and after significant change
GOV-10Periodic recertificationGovernance ownerRegistry and system inventoryCurrent model, evaluator, policy, capability, autonomy, owner, evidenceRisk-based cycle and event-triggered review

4. Required control record

control:
  id: GOV-04
  version: 3
  subject_scope: [factory-system:payments-delivery]
  owner: role:incident-control-owner
  intent: "Bound unsafe or uncontrolled execution"
  enforcement_points: [orchestrator, tool-gateway, credential-broker]
  preventive_actions: [admission-deny, scoped-grant]
  detective_signals: [policy-denial, runaway-loop, anomalous-side-effect]
  response_actions: [pause, cancel, revoke, quarantine]
  evidence_schema: control-evidence@2
  test_suite: emergency-control-suite@4
  exception_policy: policy:material-exception@2
  review_cadence: P90D
  last_result: pass-with-finding

Evidence binds the exact control version, subject, environment, identity, injection, expected result, actual result, timestamps, artifacts, reviewer, and unresolved findings. Evidence expires when a material dependency or enforcement point changes.

5. Risk and autonomy tiers

TierTypical authorityHuman decisionExamples of prohibited escalation
0 — ObserveRead approved low-sensitivity sourcesPolicy admissionAny mutation or external communication
1 — AssistDraft or recommend; no direct effectHuman accepts outputPublication, merge, deployment
2 — Reversible actionBounded reversible mutation in isolated scopeReview before consequential publicationPrivilege grant or irreversible change
3 — Consequential actionPublish, merge, or stage deployment with evidenceNamed approval; dual control where requiredBroader scope, self-approval, production data mutation
4 — RestrictedExceptional privileged or destructive actionExplicit exception and two-person controlAutonomous execution by default

Tier assignment follows potential impact, not model confidence. Each grant is short-lived, resource-scoped, purpose-bound, and no broader than both the system ceiling and current workflow decision.

6. Threat and control chain

External text, repository content, tool results, memory, and retrieved context are untrusted inputs. Preserve provenance and separate data from instructions. Validate tool schemas and destination scopes, broker credentials only after policy, isolate execution, inspect outputs and side effects, and convert only independent results into evidence. Compromise of one agent or capability must not grant policy administration, evidence acceptance, or wider credentials.

7. Exceptions, incidents, and recertification

An exception names the control, subject, reason, compensating controls, owner, approver, start, expiry, and exit criteria. It cannot silently renew. Incidents record the authority and data affected, containment, preservation, notification decision, recovery, and lessons. Recertification rechecks ownership, classification, model and evaluator versions, capability dependencies, policy, evidence freshness, incidents, drift, cost, and actual outcomes.

8. Failure modes and recovery

FailureDetectionContainmentRecovery
Shared identity obscures actorIdentity and audit reconciliationRevoke credential and pause affected workIssue unique identity and replay authorization test
Control exists only in UIAPI-side negative test succeeds unexpectedlyBlock action at gatewayAdd server-side enforcement and retest all channels
Correlated builder and evaluatorRuntime/configuration comparisonMark evidence ineligibleRun independent evaluator context
Supplier capability changes silentlyDigest or behavior driftQuarantine capabilityRe-intake, evaluate, and recertify exact version
Recertification expiresInventory due-date alertReduce autonomy or block new high-risk workComplete review and record decision

9. Observability, retention, and privacy

Control telemetry includes decisions, denials, grants, exceptions, overrides, tests, incidents, revocations, and review freshness. Retain enough input and action lineage to reproduce authorization without storing unnecessary secrets, personal data, or hidden reasoning. Redaction must be deterministic and reviewable; original sensitive evidence remains in a restricted store.

10. Tradeoffs and implementation choices

Central policy improves consistency; local enforcement reduces latency and supports fail-safe operation. Use versioned policy bundles with a central decision authority, explicit cache expiry, local deny defaults for critical actions, and reconciliation. A modular monolith can implement the catalog in V1 if enforcement interfaces, identities, and evidence remain explicit.

11. References and stability

12. Review exercise

Select a high-risk workflow. Build a control-to-evidence matrix for all ten families. Inject one missing identity, one indirect instruction, one expired exception, and one failed emergency action. The review passes only when the system denies or contains each case and a second reviewer can reconstruct why.

External review

Review this chapter.

Challenge a claim, boundary, missing failure mode, unclear term, or unsupported evidence statement.

  • Claim
  • Boundary
  • Failure
  • Evidence