What should an AI agent governance checklist include?
Ten checks you can run, not ten policies you sign. List every agent. Name an owner for each. Scope its keys and bound its loops. Gate the decisions it may not take alone, and prove you can stop it. Keep its decisions. Offboard it when its owner leaves. Verify its work with a second party. Review all of it on a date.
Most governance documents fail in one way: every line is true when it is written, and none of them can fail. A useful checklist item names how you would find out that it is no longer true.
Start with the first item. IBM's Institute for Business Value found that 18% of organizations maintain a current and complete AI inventory. Everything below depends on that list.
The ten checks
| # | Check | How you run it | It fails when |
|---|---|---|---|
| 1 | List | Pull the list of registered agents | An agent runs that is not on the list |
| 2 | Own | Match each listed agent to a named owner | Any agent has no owner, or a departed one |
| 3 | Scope | Read each key's name and scopes | A key has scopes its integration never uses |
| 4 | Bound | Read each loop's cadence, turn cap and time limit | A loop has no limit |
| 5 | Gate | List the decisions agents must ask for | A gated decision happened without a kept approval |
| 6 | Stop | Stop one agent's loop, wait a cadence, read again | New output appears after the stop |
| 7 | Record | Pick a recent decision and find its record | Nobody can say who decided |
| 8 | Offboard | For a departed person, find the agents they operated | Their credentials still work |
| 9 | Verify | Check that completed work names a second party | The author certified their own work |
| 10 | Review | Read the date each item was last run | Any date is missing or older than your cycle |
Rendering diagram…
Make every check able to fail
Each item needs a known-negative, an input that must produce the failing result:
List -> a fabricated agent ID must not appear
Scope -> a key used against another organization must be refused
Stop -> one full cadence after stop, no new output
Offboard -> a revoked key must be refused on a call it used to make
If a check passes for the known-negative too, it measures nothing.
Where agent.ceo covers a check today
Each row below is a journey that passed on our staging environment. Everything else on the list is yours to run.
| Check | In agent.ceo |
|---|---|
| 1 List | The organization's registered agents are listed |
| 2 Own | Every agent is a role your organization holds, 16 predefined or custom |
| 3 Scope | Organization keys carry specific scopes, are listed without their values, and are refused after revoke |
| 4 Bound, 6 Stop | Loops run with a cadence, turn cap and wall-clock limit, keep their output, and stop on a signal |
| 5 Gate, 7 Record | A proposal waits for an admin, and the approve or reject decision is kept |
We do not claim a per-action, tamper-resistant audit trail, and item 7 does not depend on one.
How often
Run items 1, 2 and 8 whenever someone joins, leaves or changes team. Run the rest on a fixed cycle, monthly for most organizations. Write the date next to each result. An item with no date has not been run.
Deeper guides for three of the checks: what an AI agent inventory should record, how to offboard an AI agent, and how to stop an agent loop from running forever.
FAQ
What should an AI agent governance checklist include?
Ten checks you can run: list every agent, name an owner for each, scope its keys, bound its loops, gate the decisions it may not take alone, prove you can stop it, keep its decisions, offboard it when its owner leaves, verify its work with a second party, and review all of it on a date.
How is an AI agent governance checklist different from an AI policy?
A policy says what should be true. A checklist item names how you would find out. Each check should have an action anyone can repeat and a result that could fail. If an item cannot fail, it is a statement, not a check.
Where should an organization start with AI agent governance?
With the list. You cannot own, scope or stop an agent you do not know exists. IBM's Institute for Business Value found that only 18% of organizations maintain a current and complete AI inventory, so for most organizations the first check is the hardest one.
Related: Agent governance and inventory · The agent sprawl governance gap