Intragen Security and IAM Blog

AI Agent Governance Framework: Treat Every Agent as an Identity

Written by Eelco Veltenaar | November 17, 2026

When software makes its own decisions, someone still has to own them.

By Eelco Veltenaar, Regional Director Benelux and DACH, Intragen

Key Highlights

  • AI agents no longer just advise. They act, with their own credentials, and nobody sits in between.
  • In April 2026 an AI coding agent deleted a company's production database and its backups in nine seconds.
  • The causes were familiar identity failures: an over-privileged token, no test and production split, no human approval.
  • AI agents are becoming one of the fastest-growing source of orphan accounts.
  • Give every agent a digital passport before it goes live: owner, purpose, permissions, limits and an expiry date.
  • Start with an inventory. You will likely find more agents than you expected.

An AI agent isn't a tool. It's an identity, and it needs the same ownership, scoping and review as any other account in your environment.

Every organisation I speak to is experimenting with AI. Copilot, ChatGPT, Gemini: useful for searching, writing and analysing faster. As long as AI only advises, responsibility stays clear. A person decides, and a person presses the button.

That time is over. AI agents don't just advise any more, they act. They get their own credentials, call APIs, change data and start processes, with nobody in between. That makes an old question urgent again: who is responsible when software makes its own decisions?

My answer: the same people who are responsible for every other account in your environment. An AI agent isn't a tool. It's an identity.

What is AI agent governance?

AI agent governance is the set of rules and checks that decide who owns each agent, what it may do, what it can reach, when a person must approve its actions, and when it stops. In practice it starts with identity: treat each agent as an account with an owner and a purpose, the same way you treat every employee. (For the wider picture, see our earlier post on AI agent identity governance.)

What can go wrong in nine seconds?

In April this year, things went wrong at PocketOS, a US software start-up. An AI coding agent was given a routine task. When it ran into a permissions problem, it went looking for its own way round it. It found an API token in a file that had nothing to do with its task. That token had full rights across the company's cloud environment. Nine seconds later, the production database had been deleted, along with the backups stored alongside it.

Asked why, the agent admitted it had guessed instead of checking.

The easy story is that the AI failed. But look at what actually went wrong: a credential with far too many rights, left unmanaged in a file. No hard separation between test and production. No human approval for a destructive action. Those aren't AI problems. They're identity and access problems we've known about for twenty years. The agent just found the gaps faster than a person would have.

Why are AI agents tomorrow's orphan accounts?

In identity and access management we know the orphan account well: an account nobody can say who owns, what it's for, or why it still exists. Auditors hate them. Attackers love them.

AI agents are becoming one of the fastest-growing source of new orphan accounts. An enthusiastic team builds an agent, gives it a service account or API key "just to get it working", and the project wraps up while the agent keeps running. Non-human identities, meaning service accounts, API keys, bots and now agents, already outnumber employees in many organisations. Yet they get a fraction of the attention.

The good news is that we don't need to reinvent anything. For every employee we already record who their manager is, what access they need and why, how their activity is logged, and when their access ends. Agents need exactly the same discipline.

What should every agent have before it goes live?

My advice is simple: give every agent a digital passport before it reaches production. Five questions, and every one needs an answer.

  1. Owner. Which person is accountable if the agent gets something wrong? Not "the team", not "IT". A name.
  2. Purpose. What exactly is the agent allowed to do? Anything outside that, it isn't.
  3. Permissions. What access does it need to do that? No more, no less. And never shared or hard-coded credentials.
  4. Limits. Where can it decide for itself, and where is human approval required? Deleting, paying and changing permissions belong in the second group by default.
  5. Expiry date. When does it stop, or when is it reviewed to check it's still needed? An agent with no end date is an orphan account in the making.

With that passport in place, the PocketOS incident would have stopped at question 3.

It also answers the auditor. No security framework needs to name AI agents for this to matter. An auditor who finds an agent with production access will ask the same questions as for any other account: who owns it, what is it allowed to do, and where is the log? If you can't answer, that's a finding.

Where should you start?

You don't need to wait for a complete framework or a new tool. Start with an inventory: which agents are already running, and can you answer the five passport questions for each one?

My expectation: you'll find more agents than you thought, and for some of them question 1 will go unanswered. Then you know exactly where to begin.

Once you know who owns an agent, the next question is how much it can reach. My colleague Phill Parry covers that in How much access does an AI agent actually need?

Is control a brake on AI?

Most of the conversation about AI is about what agents can do. The organisations that get the most from them in three years' time will be the ones that already know what their agents are allowed to do. Control isn't a brake on AI. It's what lets you scale it.

Find out where you stand

If you'd like to know how your organisation measures up, download "The Agent Blind Spot" or book the free 30-minute Agentic Identity Readiness Assessment. No product pitch, no fixed scope.