There is an agent-governance test that fits in one sentence and requires no tooling. Pick an agent in production and answer: what can its credential do on a Tuesday at three in the morning, with no task running and nobody watching?
Why the question works
It strips away everything that normally masks the risk. It does not ask what the agent does — it asks what it can do. It does not ask what the documentation says, it asks what IAM permits. And it removes the assumption of supervision, the favourite crutch of any loose design.
In most environments the honest answer is: everything it can do at any other hour, because the permission has no notion of time, task or context.
A permission that does not change between working and idle is standing privilege, whatever the architecture document calls it.
The three answers that define maturity
- “I do not know.” Permission inventory is missing. This is the most common stage and the easiest to fix — the information exists in IAM, nobody consolidated it.
- “Everything it does during the day.” Inventory exists, design does not. This is where per-task elevation comes in.
- “Read within scope X, and nothing else until someone requests it.” Here governance is real, and the rest of the programme gets much cheaper.
Taking this to the board
The question also works as a communication tool, because it needs no technical vocabulary. “We have eleven agents that, at three in the morning, can delete customer records with nobody approving” is a sentence any executive understands, and it produces a decision. “Our IAM posture for non-human workloads is below target” produces nothing.
The follow-up version
Once you have answered for one agent, repeat for the others and turn it into a metric: how many agents hold write permission at rest. That number should fall over time. If it is rising, the programme is losing, even if no incident has happened yet.
