Agents arrived before the paperwork did
Ask most companies three simple questions: how many AI agents are working here, who is responsible for each one, and what can they reach? Few can answer. You can't secure what you haven't found, so the first job is to find AI agents in your business. This is where AgentFort starts: finding the agents already at work.
AI agents didn't come into most companies through one big purchase. They showed up one team at a time.
The support team built an assistant in Microsoft Copilot Studio to sort incoming tickets. Sales operations switched on an agent in Salesforce. Engineering shipped one on AWS that reconciles invoices every night. And someone in finance wrote a small script that sends spreadsheets to a public AI service to summarize them.
Each one made sense to the person who built it. Most of them aren't on any list that security, IT or compliance can see.
That gap has a name: shadow AI, meaning AI in use that no platform, project or owner accounts for. It's rarely malicious. It's just invisible, and you can't protect, budget for or explain to an auditor something you don't know exists.
Why asking teams doesn't work
The usual fix is a survey or a sign-up form for new agents. It finds the agents people remember, and the ones they're comfortable telling you about.
Agents are built in many tools, by many teams, and they change every week. A list that depends on people filling it in is incomplete on day one and out of date by day thirty.
What you need instead is an AI agent inventory. An AI agent inventory is a record of every agent at work in the company: what it is, where it runs, who owns it and what it can reach. An inventory you can rely on has to come from evidence, from what agents actually leave behind, not from what people report.
What AgentFort Discover finds
AgentFort Discover sorts everything it finds into plain categories, so you know what you're looking at and where to start.
| Category | What it means |
|---|---|
| Platform agents: agents built on business platforms | Assistants and agents your teams created in Microsoft Copilot Studio, Salesforce Agentforce, AWS and Google Cloud. One list instead of four admin consoles. |
| Custom agents: agents your engineers built | Agents running in your own cloud accounts, containers and servers. Discover shows where each one runs, down to the server and the program. |
| Shadow AI: AI that nobody signed off on | Software using AI services that no platform, project or owner accounts for. It's flagged so someone can review it and take ownership. |
| AI features: AI that answers but doesn't act | A tool that drafts marketing copy uses AI but takes no action. An agent decides and acts: it sends email, updates records, issues refunds. Discover tells them apart, so you start with the ones that act. |
| In development: agents on the way | Agent projects in your code repositories that haven't been seen running yet, so you know what's coming before it reaches production. |
| Tools: the tools agents use | The services agents connect to in order to get work done, and agents that offer their services to other agents. |
How to find AI agents in your business: five places to look
Every agent depends on a few things: an AI model to decide what to do, accounts and permissions to act with, and somewhere to run. Each of those leaves traces in systems you already have. Discover looks in five places:
- Where agents are set up: the AI platforms your teams build on, such as Copilot Studio, Agentforce and AWS.
- Where they get their intelligence: the AI services they call, as recorded in your cloud's logs or seen by a sensor on your servers.
- What they can reach: the accounts and permissions they hold in your identity provider.
- Where they're built: your code repositories.
- Where they run: your cloud accounts, containers and servers.
No single place sees everything. A platform knows an agent exists, but not what else it can reach. Your identity provider knows what an account can do, but not that an agent is using it. Discover brings the pieces together into one entry per agent, with the evidence behind it, so you can see what is confirmed and what is only suspected. When it can't be sure that two pieces belong to the same agent, it asks a person instead of guessing.
What you see for every agent
For every agent, Discover answers the questions your security, IT and compliance teams ask first: what it is and how sure we are, where it runs, who owns it, and what it can reach. It also flags what needs attention:
- Shadow AI: AI in use that no platform, project or owner accounts for.
- No owner: nobody is accountable for what the agent does.
- Broad access: more access than the job needs, such as an assistant that can send email as anyone in the company.
- Shared login: one set of credentials used by several agents, so you can't tell who did what.
- No identity: the agent has no identity of its own to hold it to account.
Broad access deserves early attention on any agent that acts. OWASP lists excessive permissions as one of three root causes of excessive agency, alongside excessive functionality and excessive autonomy.
When a new agent appears, Discover tells you, and it can notify the tools your team already watches.
What a first AI agent inventory looks like
Here are four of the fourteen things Discover found in our product demo, which we recorded live against a simulated company.
| What it found | Where | What it is | Why it matters |
|---|---|---|---|
| Support triage assistant | The support team's cluster | Agent, confirmed | It can send email as anyone in the company. |
| Invoice reconciler | AWS | Agent, confirmed | The platform, the logs and the code all point to the same agent. |
| Marketing copy writer | AWS | AI feature | It writes text and takes no actions, so it's lower priority. |
| Summary script | A finance workstation | Shadow AI | It calls a public AI service, with no owner, project or approval. |
Northwind is an illustrative company used to demonstrate the discovery workflow. You can watch the 1 min 42 s demo of how Discover finds agents on the AgentFort Discover page.
No rerouting. No rebuilding.
Security leaders often ask whether they have to send all their AI traffic through us before we can find anything. You don't.
Start with what you have. Discover begins with read-only connections to systems you already use: Microsoft Copilot Studio and Entra ID, Salesforce, AWS, Google Cloud, GitHub, and the logs your cloud and applications already keep.
Nothing changes for your agents. They keep working exactly as they do today. Nothing is rerouted, and no team has to register anything.
Add detail where you want it. For agents running on your own servers, an optional lightweight sensor shows exactly where each one runs. It stays inside your environment.
Your data stays yours. Discover records that an agent exists and what it connects to. It never stores prompts, answers or the data your agents handle.
A checklist for your first AI agent inventory
Whatever tool you use, the same steps turn a guess into an inventory you can defend. For more depth, see our guide to shadow AI agents and AI agent discovery.
- Build the list from evidence, not from surveys or sign-up forms.
- Check all five places: platforms, AI services, identities, code repositories and runtime.
- Separate agents that act from AI features that only answer, and review the agents first.
- Give every agent a named owner, and flag the ones nobody claims.
- Split shared logins, so each agent's actions can be traced to that agent.
- Ask a person before merging two pieces of evidence into one agent.
- Keep the list live, with an alert when a new agent appears.
Discovery is step one
Seeing your agents is where control starts. From the inventory, AgentFort gives each agent its own identity and an owner who answers for it, limits it to what its job needs, keeps sensitive data from leaking through it, contains it if it's manipulated, and keeps a record of everything it did. Our practical guide to AI agent security covers each of those controls.
We are recruiting design partners to evaluate AgentFort in their own cloud or network. Start with a demo, or apply to become a design partner and shape what we build next.
Frequently asked questions
How do I find shadow AI in my company?
Look for calls to AI services that nothing accounts for. Your cloud's model logs, proxy logs and a sensor on hosts and pods show which programs call a model. Match each one against your AI platforms, code repositories and identity provider. Whatever has no platform, project or owner behind it is shadow AI. Flag it for review rather than blocking it outright, and give it an owner or retire it.
How can I tell whether a program is an AI agent?
A model call alone doesn't make an agent. A program becomes an agent when the model asks for tools and the tool results go back to it: an agent loop. A program that only sends text and gets text back is an AI feature. The difference matters because agents act, by sending email, updating records or issuing refunds, so they belong at the top of your review list.
Can OpenTelemetry help build an AI agent inventory?
It can add evidence. As of October 2026, OpenTelemetry maintains semantic conventions for generative AI, covering spans, metrics and events for GenAI clients, as well as MCP. If your applications emit them, traces show which services call models and tools. Traces only cover code someone instrumented, so treat them as one source next to platform connectors, identity providers, code repositories and runtime sensors, not as the whole inventory.
Who should own an AI agent?
A named person who answers for what the agent does, usually someone on the team that built or runs it, not a shared mailbox. The owner approves the agent's identity and access, reviews its risk flags and is the first call when it misbehaves. A sound rule is that an agent nobody owns gets no identity until someone claims it, so ownerless agents can't quietly keep working.