At 90 days, six metrics show whether AI agent governance is working: inventory coverage, share of agents with a named owner, reduction in excess permissions, approval quality, exception volume and age, and time to stop an agent. Choose a baseline now, because a number without a starting point proves nothing.
The six measures
| Metric | How to count it | Healthy direction |
|---|---|---|
| Inventory coverage | Agents found by two independent methods divided by agents you know about | Rising, with new finds falling |
| Owner assigned | Agents with a named owner and backup | Toward all |
| Excess permissions | Agents holding write access or broad scopes they do not use | Falling |
| Approval quality | Sample of approvals reviewed by a second person, share judged meaningful | Rising |
| Exceptions | Open exceptions and their age | Few, short lived |
| Time to stop | Minutes from decision to last action ceasing, in a test | Falling, then steady |
Why these six
They map to the root causes OWASP gives for excessive agency: functionality, permissions and autonomy. Excess permissions track the permission cause. Approval quality tracks autonomy. Coverage and ownership keep functionality visible. Time to stop addresses damage limitation, which OWASP says logging and monitoring support even when they cannot prevent the problem.
Approval quality needs a method
Counting approvals proves little, because a rubber stamp counts as an approval. Each month, pull ten approvals at random. A second reviewer asks: did the approver see the actual inputs, did the action match the stated purpose, would they approve it again? Record the share that pass. The MCP specification advises showing tool inputs to the user before calling the server, which is the standard to hold the approver's view to.
Set the baseline in week one
Write down today's numbers, even if they are poor. Coverage may be 40 percent. Half the agents may lack owners. That is fine. The story at 90 days is the change.
Avoid vanity numbers
Number of policies written, number of training sessions held and number of tools bought say little about risk. Number of blocked actions can mislead as well, because a rising count may mean more attacks, a noisier rule or more usage. Pair it with a sample review of what was blocked.
Review cadence
Look at the six measures monthly with the owners. At day 90, write a one page summary: baseline, current, what moved, what did not, and the next quarter's two goals. Show the same table every time.
Try it yourself
Click through it at your own pace, or open it full screen.
Look at how ongoing monitoring tracks change over time, the habit these six measures depend on.
Quick check
Is a lower number of agents a good result?
Not by itself. Fewer known agents may mean worse discovery. Use coverage to tell the difference.
Who should see these metrics?
The agent owners monthly and security leadership quarterly.
Safeguard's asset discovery, agent identity and policy features can supply the underlying counts, and the free scan shows a repository baseline.
Sources
Last reviewed: 19 October 2026.
