
In T. S. Eliot’s play, The Rock, the Chorus asks: “Where is the knowledge we have lost in information?” An enterprise agent inventory presents a practical version of that question: how do we turn a massive set of data into prioritized, actionable knowledge?
And where do we even start when the list is so huge?
Security teams have faced this problem before with logs, SOC alerts, and vulnerabilities. A long list of vulnerabilities doesn’t tell us what to fix first. An agent inventory needs the same work: establish what matters, why it matters, and who should act.
First Things First: Is the Inventory Complete?
Even if the inventory feels overwhelming, ask whether it covers the places agents are actually running.
That can seem counterintuitive. It’s already so big we don’t know where to start, and you’re asking if it’s big enough?
Yes. You might have 2,000 agents listed, but if the five that can reach your most sensitive data or transfer funds are missing, you could overlook your biggest risks.
Look beyond approved platforms to developer laptops, servers, cloud infrastructure, and agents embedded in SaaS applications and workflows. Include third-party agents operating in your environment.
Repositories and deployment records help establish what was built. Identity, secrets-management, and API records help establish what has access and what is being used. Work-related use through personal accounts also matters, within your organization’s authorized visibility.
Check what the inventory counts. A hundred runs of one agent aren’t necessarily a hundred agents to review. Connect those runs to the application, its owner, and its permissions.
Record coverage gaps and when each source was last checked. Keep closing those gaps while addressing the risks already visible. An incomplete inventory isn’t a reason to postpone action.
Where Do We Start When This List Is So Big?
Evidence of active misuse or harm belongs in incident response, where it can be addressed immediately rather than waiting in the inventory review queue.
For the rest, start with agents that can affect sensitive data or important systems, especially when outsiders can influence them or you don’t know what controls are in place.
In vulnerability management, a high severity rating gets our attention, but we also look at exposure, whether an issue can be exploited, and what the business stands to lose. We can bring that same judgment to the agent inventory.
Four Questions Help Establish the Priority
- What data can it reach? Customer records, financial information, sensitive documents, or secrets? Verify access rather than relying on what the agent was intended to use.
- Who or what can influence it? Who can interact with it? Can it read outside documents, web pages, emails, or tool responses? Instructions hidden in that material can redirect the agent through indirect prompt injection.
- What can it do? An agent that can read a record has different authority from one that can change or delete it, or transfer funds. Check connected tools and the permissions enforced by the systems behind them.
- What happens if it fails or acts incorrectly? Could it interrupt a critical service, expose information, harm a customer, or disrupt a business partner?

If one agent can ask another to do work, follow the request to the second agent. What can it do, and what checks does it apply before acting? The first agent may be able to cause an action through the second, even if the first agent can’t perform that action directly.
Consider two fictional customer-service agents that access the same records. One drafts replies for an employee to review and cannot send them. The other accepts customer requests and issues refunds, but its payment limits haven’t been verified.
Review the refund agent’s action controls first. There’s a direct path from an outside request to a financial transaction, with an important control unknown. The drafting agent still needs data protection; evidence that it could expose sensitive records could change its priority.
Record why an agent was prioritized so the owner knows what to address and leadership can understand the decision.
My Noma colleague Gal Moyal built on OWASP’s excessive-agency guidance in their No Excessive CAP framework: Capabilities, Autonomy, and Permissions.
The four questions establish context and consequences. CAP helps assess whether the agent’s available actions, freedom to act, and access fit its job. It provides a scoring lens to inform the decision alongside exposure, business impact, and verified controls.
In parallel, separate abandoned experiments from agents used infrequently for important work. A quarterly financial process may be quiet this month. An abandoned experiment may still hold a live credential.
Before retiring an agent, check dependencies and shared access. Remove its triggers and exposure, revoke dedicated access, and preserve required records through the normal change process. Don’t forget that deleting the inventory entry doesn’t revoke the access.
Who Owns It, and Whose Identity Does It Use?
Ownership and identity answer different questions.
The owner is accountable for the agent’s purpose and operation. A different team may maintain it. Its identity is the account it uses to access a system.
A customer-service team might own an agent that a developer maintains, while the agent acts through access delegated from that developer’s account. Knowing the business owner doesn’t tell us whose permissions apply.
Agents may use several identities across their tools, or share an account with other agents. Connect important actions to the identity used, its permissions, and the team able to change or revoke access.
If ownership is unknown, trace it through deployment records, repository maintainers, cloud-account ownership, and the business process. Assign an interim team and an investigation deadline. Immediate risks still warrant containment while ownership is resolved.
Connecting the Pieces
Pulling these details together by hand for thousands of agents won’t scale. We need enough context to make a decision, even when that context comes from several different teams and tools.
Automate collection and matching where possible. Bring purpose, ownership, identities, access, and activity into one view. Separate observed evidence from owner-supplied information, and show when each was verified. An old “internal only” label shouldn’t outweigh evidence of public access.
Trace the systems, data, and processes an agent could affect, checking controls along the way. Compare activity with its assigned job. Familiar behavior can still be unauthorized, and a quiet agent can still hold excessive access.
Keep this picture current. New tools, changed permissions, or a move into production can change the priority. Give new and materially changed agents a review path so unmanaged growth doesn’t overtake the cleanup.
Taking Action
Turn findings into work through the processes teams already use: incidents, access reviews, development tickets, and business-risk decisions. Include the evidence, business consequence, owner, next action, and deadline.
For the refund agent, that might mean enforcing transaction limits, naming who approves exceptions, and testing that requests outside those limits are stopped. A model’s recommendation to refund a customer doesn’t authorize the payment.
Define how to verify completion. Test that unwanted access has been removed. Confirm a retired agent has stopped running and its dedicated credentials are revoked. If the business accepts remaining risk, record the authorized decision-maker, rationale, and review date.
Measure those outcomes. A growing inventory may mean discovery is improving; the count alone says little about risk.
Start with a handful of agents with consequential access or authority. Establish what each can do, identify its accountable owner, and agree on an action and evidence of completion.
We’ve got a lot of information. Connecting it to consequences and decisions is how we turn it into knowledge.





