In short
AI agent governance is the set of policies, roles and controls determining who may create AI agents, what data and tools each agent may use, how its behaviour is reviewed, and when it is retired. Effective governance is enforced in the platform rather than documented in a policy, and it covers the AI Principals that create agents as well as the agents themselves.
Definition
AI agent governance answers ownership questions: who authorized this agent, what business purpose justifies its access, who reviews its behaviour, and what happens when the owner leaves. It sits between AI governance, which addresses model risk, fairness and compliance, and identity governance, which addresses credentials and entitlements.
The governing insight is that agent populations grow at machine speed while approval processes move at audit-cycle speed. Governance that depends on tickets and spreadsheets is stale on arrival. The workable model is policy-as-code enforced at creation and at every action, with periodic human review of exceptions rather than of every instance.
NIST AI RMF provides the organizational structure — govern, map, measure, manage — and ISO/IEC 42001 supplies a certifiable management-system framing. Both are compatible with the technical controls described in OWASP's agentic guidance.
Security risks
Governance on paper only
Policies that no system enforces produce audit findings, not risk reduction.
No accountable owner
Orphaned agents keep valid access indefinitely because no leaver process covers them.
Approval bottlenecks
Slow review pushes teams to build outside the process, converting governed demand into shadow AI.
Uncontrolled creation authority
If any team or agent can spawn agents, population growth outpaces every downstream control.
Unmeasured behaviour
Without metrics on action volume, privilege use and exceptions, governance cannot tell healthy from harmful.
Compliance exposure
Regulated processing performed by undocumented agents is difficult to defend at audit or under the EU AI Act.
Examples
Agent registration at creation
The orchestration platform refuses to spawn an agent without an owner, purpose, data classification and expiry date.
Tiered approval
Low-impact agents self-serve; agents touching regulated data or holding write access require security review.
Quarterly access recertification
Owners confirm each agent is still needed and still correctly scoped; unconfirmed agents are disabled automatically.
Leaver-triggered revocation
When an owner departs, their agents are reassigned within a fixed window or automatically suspended.
Architecture
Agent register
Authoritative record of every agent: owner, purpose, data scope, tools, model version, creating Principal and expiry.
Policy-as-code
Machine-readable rules evaluated at creation and at each action, versioned like any other code.
Creation gateway
The single path through which agents may be spawned, applying policy before the agent exists.
Review and recertification workflow
Scheduled owner attestation with automatic consequences for non-response.
Metrics and reporting
Population size, privilege distribution, exception counts, orphan rate and time-to-revoke.
Controls
Make ownership mandatory at creation
No owner, no agent. This single rule prevents most orphan and offboarding failures.
Enforce policy in the platform
Controls must live where agents are created and where tool calls execute, not in a policy document.
Govern creation authority explicitly
Maintain a register of AI Principals, restrict who may hold creation rights, and rate-limit spawning.
Set expiry by default
Every agent and credential expires unless renewed with justification.
Provide a fast sanctioned path
Make the compliant route quicker than the workaround, or governance will be routed around.
Measure the programme
Track orphan rate, over-privileged agent count, mean time to revoke and percentage of agents with full traces.
Comparison
| AI governance | Identity governance | AI agent governance | |
|---|---|---|---|
| Primary object | Models and use cases | Users and credentials | Agents and their creators |
| Typical cadence | Project and release | Quarterly recertification | Continuous, at creation and action |
| Key artifact | Model risk assessment | Entitlement review | Agent register with lineage |
| Enforcement point | Approval process | IAM system | Creation gateway and tool broker |
| Main failure | Shelfware policy | Rubber-stamped reviews | Population growing faster than review |
Frequently asked questions
Who should own AI agent governance?
Usually security in partnership with the platform team that operates the agent runtime, with business owners accountable for individual agents.
How does this map to the EU AI Act?
Obligations depend on use case and risk tier, but a complete agent register, traces and human oversight are prerequisites for demonstrating compliance in any tier.
Do we need ISO/IEC 42001?
Not universally, but it provides a recognized management-system structure and is increasingly requested in enterprise procurement.
What is the first policy to write?
A creation policy: who may create agents, what metadata is mandatory, what privileges are permitted by default and when agents expire.
Sources
How Andeavour addresses it
Enforced register, not a spreadsheet
AgentGuard maintains a live inventory of agents and Principals with owners, scope and lineage, updated continuously rather than at review time.
Policy at creation and at action
Rules are applied when agents are spawned and again when they call tools, so governance keeps pace with machine-speed growth.
Audit-ready evidence
Traces, approvals and revocations produce the documentation required for ISO/IEC 42001, NIST AI RMF and customer assurance.
See it against your own environment
A 30-minute technical session with the Andeavour team — no integration required to start.