By Keith Rull ·
Give an AI agent a credential only when one named person answers for what it does and a date ends its access. If you cannot name both, leave it off. The person is who you call when it does something wrong. The date is what turns the credential off after the project ends.
Why do agents outlast the work that created them?
Nearly twenty years ago I was a junior developer on an SMS ordering app for a pharmaceutical company. One night the app stopped sending and receiving texts. Orders still landed in the database and sat in a queue. Nothing went out, and nothing came back.
I restarted the application, the services, and the server, then walked the code for about ten hours. The only signal was a DLL exception, an error from a Windows library, and the null reference did not point at the cause.
At 6 a.m., with about two hours left before I had to face the customer, I stepped out of the server room and told the janitor. He had vacuumed that room the day before, could not find a free socket, and unplugged something near the server without being sure he had plugged it back in. It was the GSM modem. Once it was plugged in, the texts started moving again.
He had done this more than once, and nobody caught it, because he usually cleaned late at night and that day he cleaned early. He was doing his job. The failure was unwritten access to production: I had spent ten hours in the code, and the useful question was who else could change the system and what they had changed.
An agent left running after a project ends has that same shape: it is still following its instructions, the task is over, and the credential is not. When an employee leaves, an HR record can start the removal of their access, and an agent has no equivalent last day. ServiceNow’s 2026 report on agentic security makes the identity point: nothing in the environment notices that the project is over, so the credential keeps working. The report is Deployed but Ungoverned: The State of Agentic Security.
There is a quieter version of the same gap. Someone approves access in a ticket and writes “temporary, through the end of the quarter,” the quarter ends, the ticket says expired, and the identity system never received a date, so the access is still live. A date that lives only in a document did not end anything.
The agent you reviewed on day one is also a different agent on day ninety, because prompts, tools, and the data it can touch all change. Permission granted once does not stay correct on its own.
What do the surveys actually show?
Two 2026 vendor surveys describe the gap. Both are self-reported, so treat them as a floor rather than an audit of your company.
ServiceNow, with Phronesis Partners, surveyed 3,766 VP-level security and technology leaders across 20 countries. In the previous 12 months, 29% of security teams found unregistered AI agents in production, and only 34% register a new agent automatically at deployment. The rest depend on someone remembering. And 66% cannot act automatically when an agent behaves unexpectedly, so a person has to notice first.
The same report says 76% of leaders keep a complete, current map of every agent’s access, while only 19% can see more than 90% of their agents in real time. That gap is a list versus what is actually running.
Gravitee’s October 2026 wave asked 900 senior technology leaders at organizations of 250 or more employees in the UK and the US, with fieldwork from 8 to 19 October 2026. It found that 9.5% had a named individual formally accountable for agent behavior, up from 7.2% in April 2026, which still leaves about nine in ten with no named owner. Only 6.3% could revoke a misbehaving agent’s access to all systems within an hour, and 46% would need more than a day. The write-up is the State of AI Agent Security.
What do a name and an end date require?
One person, not a team name
A name means one human who answers for the outcome, which a team name, the word “platform,” or “whoever built it” cannot do. You want one person who can explain what the agent was supposed to do, and who gets the call when it does something wrong.
That person is separate from the off switch, and you need both. The owner answers for the outcome, and the off switch has to work when the owner is on a plane. If they are the same person and there is no backup, you have a single point of failure.
An end date the identity system enforces
Treat each agent like a contractor. A contractor gets a start, an end, and a manager who has to renew them. If nobody renews, access ends. That flips the default from “runs until someone remembers” to “stops unless someone vouches for it.”
If the record says 90 days, the credential should stop working on day 91 with no one clicking a button. Renewal should be a short review of whether the scope, the tools, or the data has changed. A one-click renewal that nobody reads is a rubber stamp.
An inventory that starts from credentials
Skip the ask for teams to list their own agents. You will collect the optimistic number. Start from the other direction: every API key, service account, and OAuth grant. A credential with no owner and no expiry is your unregistered agent. Then go find out what it does.
Seeing the credential is not the same as knowing the agent finishes the task. I look at outcome, failed tool calls, and whether a person can recover the handoff in AI agent reliability.
This register is hypothetical. None of these rows is a real system. It is the one-page list I would want before a migration agent kept its key.
| Credential | Owner | Expires | What it can reach | Next step |
|---|---|---|---|---|
| API key on the migration job | Platform lead | 2026-12-15 | Read the old bucket, write the new one | Renew or disable |
| OAuth app “inbox sort” | Unassigned | None | Send mail as a shared mailbox | Disable, then see who calls |
| CI token for a test agent | Build owner | 2026-11-30 | Open pull requests in one repo | Still inside its window |
The second row is the one I would turn off first. No owner and no date means it should not be running.
What would I do in the first week?
Companies already do this for people. The same steps cover workers that do not go home, and they do not need a new platform.
- Pull every non-human credential: API keys, service accounts, OAuth apps, and tokens in CI. That list is closer to the real inventory than a spreadsheet filled in from memory.
- Put an owner and an expiry on each one. If a credential still has no owner after two weeks, disable it and see who complains. The person who complains just volunteered.
- Put the expiry in the identity system, not only in the ticket, so access that says 90 days fails on day 91 by itself.
- Make renewal a real look at scope, tools, and data. It should take a minute of thought.
- Test the off switch on a quiet day. Time how long it takes and write down what breaks. If you cannot do that calmly on a Tuesday, you will not do it quickly during an incident.
A small team with a handful of keys can run this list. Ask for help when the list is long, nobody agrees who owns the off switch, or an agent already touches production data and the review keeps slipping. If that agent is already failing on real work, start with AI project rescue before you widen what it can reach. Pilots often stall before anyone asks who owns the agent, which is the earlier break in why AI pilots stall before production.
Put the same owner and expiry into the build, not only into a policy memo. AI development treats access and rollback as part of the release. A credential you cannot revoke is not a reversible change, which is the test in how to keep an architecture decision reversible. When the agent sits between two teams and nobody will name the owner, that is an embedded leadership problem: an embedded CTO is there to put the decision with the people who run the workflow.
FAQ
Should an agent ever hold a production credential?
Yes, when the task needs it: one person owns the outcome, the permission is as small as the task, actions outside your systems are logged, and the access expires on its own. If you cannot name the owner and the end date, it should not hold the credential.
Is a shared channel enough of an owner?
A channel does not answer the phone, and it does not renew access. Name one person, and name a backup who can turn the agent off when that person is away. The backup does not replace the owner. The backup makes the off switch work on a bad day.
What if the agent is only a proof of concept?
Give it the same two fields, and give it a short date. Proofs of concept are where forgotten credentials come from, and a demo that still works next quarter is how an unregistered agent ends up in production.
Who sets the rule, security or the team that built it?
Both, with one named owner for each agent. Security can require the expiry and the log. The person who answers for the outcome has to sit close enough to the work to notice when the scope changed. A policy that the building team cannot name will not get renewed.
First published on LinkedIn, Nov 19, 2026: Every AI Agent Needs a Name and an End Date.
If an agent still holds a credential and nobody owns it, talk to us.