Audit Your Agent Fleet Before Review
Run a pre-review agent audit, count only generally available controls, and treat missing task-scoped grants and delegation tracing as gaps to close.
The review will ask for controls, not intentions
Your agent fleet is on the review agenda. Gartner projected that the average global Fortune 500 enterprise would use more than 150,000 AI agents by 2028. Gartner also reported that only 13% of organizations believed they had the right AI agent governance in place. The Blueprint Alliance, a vendor group for agent governance, was announced on September 22, 2026, at Oktane 2026. Its reference architecture asks four questions: locate agents, define permitted actions, observe runtime behavior, and decide response. The gap is audit discipline.
The four questions are the audit
Locating agents is the first failure point. A fleet can include approved agents, experimental agents, and agents created by developers who never registered them. The audit needs a source of truth, not a spreadsheet that updates rarely. Define permitted actions next. A permission that outlives the task becomes standing access. Observe runtime behavior after that. A policy on paper does not show what an agent actually calls. Decide response last. A slow stop leaves the incident response plan a draft.
Agent identity and access scoping belong in the same pass. An agent without an identity is hard to revoke. An agent with broad access is hard to defend. Least privilege is the default. Grant only the tool the task needs, only the dataset the task needs, and revoke the grant when the task ends.
The checklist turns principles into tests
The architecture rests on six principles: agents are treated as first-class identities, permissions are limited to the task, delegation remains traceable, runtime behavior is monitored continuously, containment is immediate and reversible, and governance keeps pace with AI. Run each test before the review. Record the gap and owner for every no.
- Agent identity. Locate every agent, its owner, and its identity record.
- Task-scoped access. Define permitted actions and confirm each grant is limited to the task.
- Traceable delegation. Verify every delegated action traces to the authorizing agent or user.
- Runtime monitoring. Confirm tool calls and delegated actions are monitored continuously.
- Reversible containment. Decide response and test immediate containment and rollback.
- Governance cadence. Check that governance updates keep pace with agent changes.
- GA versus roadmap. Mark each control as generally available or planned; count only GA as current.
GA is the only control you can count
Okta's status split shows why GA matters. At the announcement, Okta said Agent SSO, Agent-to-Agent Connections, and Resource Access Certifications were generally available; Agent Gateway and Shadow AI Agent Discovery for Endpoints were planned for the third quarter; Configuration Designer and the gateway kill switch were planned for the fourth quarter.
Okta positioned Agent Gateway as an inline runtime control point that enforces policy and records interactions between an agent and the tools it invokes. Use it as the test for runtime monitoring. Record the gap and add tool-call logging first when the stack lacks an inline control point.
GA status changes the answer to a reviewer. A generally available control can be cited as a current control. A planned control can be cited as a mitigation, with a date and an owner. A mitigation without a date is a wish. A date without an owner is a rumor.
Runtime monitoring is where policy meets behavior. Logs should show the agent, the tool, the action, the result, and the time. Logs must answer who acted and what changed. A log that cannot answer is not evidence. Keep the logs in a place the reviewer can query without asking a developer for a favor.
Containment should be immediate and reversible. Immediate means the stop works before the next tool call. Reversible means the team can restore a safe state without deleting evidence. A human in a hallway is a delay to document. A script must be tested.
Governance that keeps pace with AI needs a cadence. A scheduled review is a floor, not a ceiling. New agents, new tools, and new permissions should trigger a check. Fleet growth faster than policy means the policy is lagging. Policy changes without a test are promises.
Count GA capabilities as current controls. Treat roadmap items as mitigations to schedule, not evidence. Close gaps directly: reduce standing permissions, add delegation logs and approvals, start with tool-call logs, and define a manual stop and rollback.
An AI agent incident response plan should name the stop, the owner, and the rollback. Tool governance belongs in the same audit when agents call external services. Discovery matters when inventory is incomplete. The review needs a defensible gap list.
Before the review, produce an agent inventory, a gap list, and a response plan. The inventory shows what exists, the gap list shows what is missing, and the response plan names who acts when something goes wrong. A reviewer can argue with a gap. Keep the test result with the gap list.