Agentic AI Governance

Agentic systems need explicit identity, authority envelopes, stop-anywhere mediation, bounded memory, endpoint scope, evidence, repair, revocation, and exit. Autonomy is not ambient permission.

Plain-English summary

An agent should never inherit authority merely because it can reason, plan, call tools, or preserve memory. Every externally meaningful action should be traceable to a disclosed actor, bounded purpose, permitted tool, scoped endpoint, evidence record, and revocation path.

Five-plane containment model

Identity plane

Names the actor, controlling party, automation level, workload identity, and revocation owner.

Policy plane

Defines allowed tools, resources, data classes, budgets, purposes, escalation rules, and forbidden actions.

Execution plane

Mediates each tool call or endpoint action and can stop, refuse, narrow, or pause at every step.

Observation plane

Records bounded action evidence without turning all activity into behavioral surveillance.

Evidence and repair plane

Preserves receipts, errors, appeals, corrections, revocations, and exit evidence.

Intent, Plan, Observe, Evaluate

  1. IntentState the participant-approved purpose and expected outcome.
  2. PlanList tools, endpoints, data classes, budgets, dependencies, and approval triggers before acting.
  3. ObserveRecord bounded evidence about actions, denials, errors, and side effects.
  4. EvaluateCompare the result with intent, scope, evidence, and claim boundaries before continuing.

IPOE is presented here as a review pattern, not as a VNWO runtime orchestrator.

Stop-anywhere mediation

A bounded agent pipeline should be able to stop before planning, before tool selection, before credential use, before endpoint execution, after an unexpected result, before publication, and during exit. A stop should preserve a reason code and leave the system in a reviewable state.

Example stop points and evidence
Stop pointTriggerEvidence
Before tool callTool not in authority envelopeNO_AUTHORIZATION or SCOPE_EXCEEDED receipt
Before sensitive actionHuman approval triggerApproval request with purpose and scope
After unexpected resultMaterial deviation or errorIncident record and repair path
Before external publicationExternal effect not pre-authorizedHuman review and release receipt

Agent authority receipt

An authority receipt connects declared intent to the actual plan, tools, endpoint scope, memory classes, budgets, approval triggers, evidence, revocation, and exit conditions.

Authority receipt fields
Intent and outcomeParticipant-approved objective and expected deliverable.
Plan and toolsPermitted steps, tool allow-list, endpoint scope, and data classes.
Budgets and approvalsTime, compute, calls, writes, publication, credential, and human-review triggers.
Evidence and repairAction receipts, refusals, errors, side effects, correction path, and appeal owner.
Revocation and exitHow tools, endpoints, memory, credentials, delegations, and retained evidence are ended or exported.

Open the local agent and file-memory worked case.

Copyable agent authority envelope

{
  "artifact_type": "agent_authority_envelope",
  "schema_version": "3.0.8",
  "created_utc": "2026-06-21T02:42:50Z",
  "agent_id": "example-agent",
  "controlling_party": "example-organization",
  "purpose": "review public documents and prepare a draft",
  "allowed_tools": ["public_web_read", "local_draft_write"],
  "prohibited_tools": ["credential_access", "private_network_probe", "external_publish"],
  "data_classes": ["public", "project-local"],
  "budgets": {"requests": 50, "runtime_minutes": 20, "external_effects": 0},
  "human_approval_triggers": ["external publication", "new endpoint", "scope expansion", "sensitive data"],
  "stop_anywhere": true,
  "revocation_path": "/endpoint-boundaries/",
  "appeal_and_repair_path": "/repair-appeals/",
  "exit_path": "/exit-rights/",
  "claim_boundary_notes": "VNWO documents the authority envelope; it does not execute or authorize the agent."
}

Agentic governance QA checklist

  • The agent is disclosed and linked to a responsible controlling party.
  • Authority is expressed as an explicit envelope rather than inferred from capability.
  • Tool calls and endpoint access are mediated independently.
  • Memory use is scoped and does not silently become public identity or permanent context.
  • Human approval triggers and stop points are testable.
  • Revocation, repair, appeal, and exit disable residual authority and account for retained evidence.

Build networks people can leave and still choose to trust.