# AI Agent Discovery: How to Find and Account for Every Agent

> You can't govern an agent you haven't found. What AI agent discovery covers, the five kinds of agent it has to find, what to record for each one, where to look, and how to turn a list of agents into a check every agent has to pass.

Markdown version of <https://www.hawcx.com/blog/ai-agent-discovery>. Published 2026-11-19. Industry trends.

AI agent discovery is finding every AI agent that can act in your company, registered or not, and recording three things about each one: who it acts for, what it can reach, and what credentials it holds. Without them, you can't decide what it may do.

This guide covers why discovery matters, the kinds of agent it has to find, what to record, where to look, and how to turn the list into control.

## Why AI agent discovery matters

An AI agent is software that takes actions with someone's access: it reads mail, changes records, opens tickets, moves files. Agents arrive from every direction. Teams build them on purpose, employees connect assistants to their own accounts, and software vendors switch on agent features inside tools you already pay for.

Each of those agents is an identity that acts, but most don't look like one. They borrow access through an OAuth grant, an API key, a service account or a person's session, so your logs show the person or the account, not the agent. That gap is why the OWASP Top 10 for Agentic Applications lists [Identity and Privilege Abuse (ASI03) and Rogue Agents (ASI10)](https://genai.owasp.org/2025/12/09/owasp-top-10-for-agentic-applications-the-benchmark-for-agent) among the main risks of agentic systems.

![Example: one person, Dana, with four agents on borrowed access (an OAuth grant, an API key, a service account and a personal token). The audit log shows Dana and a service account, never the agents.](https://www.hawcx.com/assets/blog/ai-agent-discovery/borrowed-access.svg)

The cost of not knowing shows up in breaches. IBM's 2025 Cost of a Data Breach report found that a high level of shadow AI [added USD 670,000 to the global average breach cost](https://www.ibm.com/think/x-force/2025-cost-of-a-data-breach-navigating-ai). The same report found that [97% of organisations with an AI-related security incident](https://www.ibm.com/think/x-force/2025-cost-of-a-data-breach-navigating-ai) lacked proper AI access controls.

Every control you might want, from limits per agent to human approval to a way to stop one, starts with knowing the agent exists.

## The five kinds of agent discovery has to find

Discovery is not one search. Agents live in different places depending on who made them.

| Kind of agent | Who makes it | Where it usually lives |
| --- | --- | --- |
| Agents your teams build | Engineering, IT and business teams, on purpose | Your cloud, agent platforms, CI/CD pipelines |
| Assistants employees connect | Individual employees, for their own work | OAuth grants on their accounts, browser extensions |
| Agents inside SaaS tools | A vendor feature, switched on by a team admin | The SaaS tool's own admin console |
| Local agents and tool servers | Developers and power users | Laptops: coding tools, MCP server configs, scripts |
| Leftovers | Pilots and prototypes nobody switched off | Service accounts and keys that still work |

The first kind is usually known. The other four are where the surprises are: agents that work, that someone relies on, and that nobody reviewed.

## What to record for every agent

A useful inventory is short. Record the fields that decide what happens next, and skip the rest.

| Field | Why it matters |
| --- | --- |
| The person it acts for | Someone has to answer for what the agent does. No person means the agent is orphaned. |
| Purpose | What it is for, in one line. It tells you what it should and shouldn't need. |
| Systems and tools it can reach | Its blast radius if it misbehaves or is hijacked. |
| Sensitive actions | Anything that can move money, change access or delete data. |
| Credentials it holds | A standing key or long-lived token is what an attacker takes. |
| How it authenticates | Its own identity, or borrowed access (a person's grant, a shared key). |
| First and last seen | Agents that haven't acted in months are candidates to retire. |
| Status | Governed (enrolled, with its own identity) or ungoverned. |

If one field matters most, it's the person. An agent tied to a person can be reviewed, limited and switched off with that person's role. An agent tied to nobody can only be found again later.

Here is what that looks like for one agent:

![Example of what discovery should show for one agent, invoice-matcher, found March 3: it acts for Priya in Finance, reaches the ERP (write) and a shared inbox, holds an API key 214 days old and signs in as a shared service account. The key and the shared account lead to the decision: replace its access with its own identity, derived from Priya.](https://www.hawcx.com/assets/blog/ai-agent-discovery/one-agent-recorded.svg)

## Where to look

Agents leave traces wherever they got their access. Discovery looks in each of these places, because no single one shows every agent.

- **Consent grants in your identity provider.** Apps and assistants that people let into their mail, files and calendars. In Microsoft Entra, for example, each enterprise app's Permissions page shows [what the organisation and individual users consented to](https://learn.microsoft.com/en-us/entra/identity/enterprise-apps/manage-application-permissions), and Microsoft notes that revoking a grant doesn't stop users from consenting again.
- **Workspace admin consoles.** In Google Workspace, for example, the API controls page lists the [third-party apps people have used to reach company data](https://knowledge.workspace.google.com/admin/apps/control-which-third-party-and-internal-apps-access-google-workspace-data), each of which can be marked Trusted, Limited or Blocked.
- **Keys, secrets and service accounts.** Credentials with no clear owner, created for a pilot, or never rotated.
- **SaaS admin consoles.** Agent and automation features a team admin enabled, and what they can reach.
- **Agent platforms and pipelines.** The places your teams build and deploy agents, and the identities they run as.
- **Traffic to model providers and tool servers.** Calls from hosts or accounts that aren't on your list.
- **Managed devices.** Local agents, tool-server configs and tokens in config files.
- **Spend.** Subscriptions to AI tools and agent platforms on corporate cards.

Expect the first count to be higher than anyone guessed, and expect each source to find agents the others miss.

## Discovery is continuous, not a project

An inventory is a picture of one day. By next week someone has connected a new assistant, a vendor has shipped an agent feature, and a pilot has quietly become production. A list finished once starts going stale immediately.

Treat discovery as a running process with a clear trigger for review: a new consent with broad scopes, a new key on a sensitive system, a new destination in your traffic, an agent nobody claims. Each of those should land in someone's queue, not in a quarterly report.

## From discovery to control

Finding an agent is the start. For each one, the inventory fields decide what happens next.

| What you found | What to do |
| --- | --- |
| Useful, and a person answers for it | **Enrol it.** Give it its own identity derived from that person, and limit it to what its job needs. |
| Useful, but on a shared key or someone else's account | **Replace its access.** Move it off the borrowed credential. |
| A leftover, or nobody answers for it | **Retire it.** Revoke its grants and keys, and confirm it stops. |
| Unknown, and touching sensitive systems | **Deny it** until a person claims it and it is enrolled. |

Then make the result hold. The durable control is a check every agent has to pass, not a longer list. Put a point between agents and the systems that matter, such as a gateway, your network egress or a managed device, and deny any agent there that has no identity of its own. A new agent then fails on its first call instead of waiting for the next review.

![Three enrolled agents, each acting for a person, are allowed through a point you control to sensitive systems. An unknown script with no identity is denied at that point.](https://www.hawcx.com/assets/blog/ai-agent-discovery/the-check.svg)

Keep two limits in mind. A check covers only the traffic that passes through it, so route sensitive systems through it first. And deny by default where the stakes are high first, while people enrol the agents they rely on, or they will route around you.

For the identity half of this, see [why OAuth isn't enough for AI agents](https://www.hawcx.com/blog/why-oauth-isn-t-enough-for-ai-agents) and [agent identity rooted in an employee](https://www.hawcx.com/ai-agents/identity).

## How Hawcx approaches AI agent discovery

Hawcx Discovery finds the agents that can reach your systems and flags each one governed or ungoverned. Governed agents have an identity rooted in an employee. Ungoverned agents are enrolled, or denied by default.

Every agent, governed or not, then passes the same check. Before a tool call executes, the Hawcx Authorizer checks two things: a valid identity rooted in an employee, and an action within this agent's ceiling, the most your policy lets it do. Each call ends in one of four outcomes: allowed, held for approval, rejected, denied. An agent with no identity is denied.

Each agent's identity is derived from the employee it works for. That person signs in with Hawcx passwordless or your identity provider, such as Okta or Microsoft Entra ID.

Hawcx can check only the traffic that passes a point you control: the Hawcx Gateway, your network egress or a managed device. Anything without a valid identity is denied there.

See how this works across the [AI agent authentication platform](https://www.hawcx.com/ai-agents), or for [agents nobody registered](https://www.hawcx.com/solutions/shadow-ai-agents).

## Frequently asked questions

### What is AI agent discovery?

AI agent discovery is finding every AI agent that can act in your company, registered or not, and recording who it acts for, what it can reach and what credentials it holds. It is the first step before you can limit, approve or stop any agent.

### What is the difference between AI agent discovery and an agent inventory?

Discovery is the ongoing work of finding agents. The inventory is what it produces: one record per agent. An inventory without continuous discovery goes stale within weeks, because new agents appear faster than reviews happen.

### What should an AI agent inventory include?

At minimum: the person the agent acts for, its purpose, the systems it can reach, its sensitive actions, the credentials it holds, how it authenticates, when it was first and last seen, and whether it is governed.

### How often should we run AI agent discovery?

Continuously. New agents appear whenever someone grants a consent, creates a key or turns on a SaaS feature. Review on triggers, such as a broad new consent or an agent nobody claims, rather than on a fixed calendar.

### What should we do with agents nobody owns?

Treat them as orphaned. Revoke their grants and keys, confirm they stop, and keep a record of what they could reach. If someone claims one later, enrol it with its own identity derived from that person instead of restoring the old access.
