To find AI agents already running in your company, check six places: developer machine configs, code repositories, identity and sign-in logs, cloud accounts, expense records, and a short survey of team leads. No single source catches everything, so combine them and mark each finding with how you learned of it.
Where agents leave traces
| Source | What to look for | Catches |
|---|---|---|
| Developer machines | Tool server config files, assistant extensions | Local assistants and MCP servers |
| Repositories | Config files for agents, workflow files that call model APIs | Agents wired into CI and code |
| Identity logs | OAuth grants to AI apps, new service accounts | Connected assistants and bots |
| Cloud accounts | Functions and services calling model APIs, new API keys | Production agents |
| Expense and billing | Subscriptions and API charges for AI services | Tools bought outside IT |
| Team survey | "What do you use, and what can it touch?" | Everything else, plus context |
Start with what is cheapest
Begin with the survey and the repositories. They take a day and show you the shape of the problem. Ask the survey questions in a neutral tone and say the goal is a list, not punishment. If people fear a ban, they will not answer.
Searching repositories is concrete. Look for MCP configuration files, agent instruction files, and calls to model provider SDKs. Each hit gives an owner through commit history.
Then look at access
For every agent you find, ask what it can reach. The question matters more than the name. An agent with a read-only token to a docs site is a small matter. One holding a deploy key is not. OWASP lists excessive permissions as one root cause of damaging agent behaviour, so capture permissions in the first pass instead of after the list is done.
Record a minimum entry
For each agent, write down:
- Name and what it does
- Owner
- Tools and servers it uses
- Data and credentials it can reach
- Where it runs
- How you found it
Our post on registry entry fields goes deeper.
Keep it current
A list found once goes stale in weeks. Repeat the repository and identity checks monthly, or automate them, and compare with the last run. New entries get an owner within a week.
Try it yourself
Click through it at your own pace, or open it full screen.
Look at how an inventory is built from what is actually in a codebase, which is the same idea applied to agents and their libraries.
Safeguard provides asset discovery and agent identity as GA capabilities, and the free scan is a way to see an inventory of one repository.
Quick check
What if teams will not admit to using agents?
Lean on technical sources first, then talk about findings as a path to getting the tool approved instead of removed.
Do browser chat tools count?
For governance of agents, focus on tools that can act or reach company systems. Chat only tools are a data handling question, covered by your acceptable use rules.
Sources
Last reviewed: 19 October 2026.
