Twelve questions for your EDR vendor about AI agents

Why ask your EDR vendor first?

EDR is on every managed device already, and it watches processes, files and network activity. It was built to find malware and intrusion, not to judge whether a trusted AI agent run by a real employee should install a package. But several AI endpoint products use EDR as their delivery path: Bay and Noma Security state they deploy through existing EDR or MDM, Koi (Palo Alto Networks) is offered as a Cortex XDR module, and Prompt Security is built into SentinelOne Singularity. Your EDR vendor may have an AI answer of its own, or be the route another vendor's answer takes.

What should you ask about visibility?

1. Which AI agents, coding assistants and desktop AI apps can you identify on our devices today?
2. Can you list the MCP servers configured for each agent, including ones a user added?
3. Can you see AI browser extensions and which browsers they are in?
4. How quickly does a newly installed agent or MCP server appear?

What should you ask about control?

1. Can you stop a specific action by a trusted AI agent, such as a package install or a cloud CLI command, without blocking the agent entirely?
2. Can you ask the user to confirm instead of blocking?
3. Can you run a rule in a watch-only mode before enforcing it?
4. Does your record separate what a person did from what an agent did on their behalf?

What should you ask about delivery and roadmap?

1. Can you deploy another vendor's AI control to our devices, and what does that involve?
2. Do you offer an AI agent module or add-on, and what does it cover today rather than on the roadmap?
3. What do you see inside virtual desktops and cloud development environments?
4. What, if anything, do you offer for devices we do not manage?

How do you read the answers?

If your EDR vendor sees most AI agents and MCP servers and can act on individual agent actions, the gap is small, and a dedicated product needs to add something specific, such as an ask response or a person-versus-agent record. If it sees agents but cannot act on their actions, the gap is in decisions, which is where on-device AI controls focus. If it sees little, the gap is wide.

Pay attention to the difference between today and the roadmap. Several large security vendors have bought AI security companies in the past year, and plans are easy to describe. Ask for what is available in the version you would deploy this quarter.

What if the answers are vague?

Ask for a demonstration on one of your own developer machines with a coding agent and an MCP server installed. Watching what the console shows is quicker than reading a feature list, and it shows the difference between seeing an agent and understanding what it does. If the vendor cannot do that this quarter, note it and move on to the next question.

What about device management?

Ask your device management team the same visibility questions, plus one more: which AI agents offer administrator-managed settings, such as permission rules and allowed MCP servers, and can we push them? Bay describes managed settings for Claude Code, Codex and Claude Desktop that lock permission rules, preserve managed hooks, restrict MCP servers and constrain plugin sources. Where an agent's vendor documents such settings, device management can apply them even before any AI control is bought.

What comes after the answers?

Put them next to the finder. In the stack question, pick the platforms you run; the finder gives points to products built into them or working with them, and switches on a check for products that roll out through EDR or device management. The answers from your vendors tell you how much weight to give that result.

Who should be in the room?

The person who runs EDR, the person who runs device management, and someone from engineering who uses coding agents every day. The first two know what the tools can see; the third knows what the agents actually do and which controls would get in the way. Write the answers down in one place, with the date, because they will change. Vendors add AI features quickly, and an answer of no this quarter may be yes by the next renewal, which is a good moment to ask the twelve questions again.