AI agent control plane: what it is and what to look for
Agents do the work. A control plane decides what they are allowed to do while they do it. It holds the policy, the approvals and the record, and it is kept apart from the agents themselves.
What a control plane is
In networking and in Kubernetes, the control plane is the part that decides and the data plane is the part that does. For AI agents the split is the same. The agents and their tools are the data plane: they read, write and send. The control plane sits on the path between them and answers one question for every tool call: does this run, wait for a person, or get refused?
It has to sit outside the agent. An instruction in a prompt is a request to the model, and a model can be talked out of it by the next thing it reads. A check on the path of the call cannot be talked out of anything.
What it should do
- Identity: each agent connects as itself, has its own level, and can be stopped on its own.
- Policy by action type: decide by what a call does (read, write, share, send, delete, change settings), not by the tool's name or what its description claims.
- Approval: hold the risky calls for a person, show the whole request, and run exactly what was approved, once.
- Autonomy you can set: more room for the agents and actions you trust, less for the rest.
- Keys: hold the credentials for the tools, so the agents never do.
- Record: one tamper-evident entry for every decision, refusals included.
- A stop: one control that cuts an agent off at once.
What it is not
It is not an agent framework: it does not plan or run your agents. It is not an observability tool: seeing a call after it ran is different from deciding before it runs. And it is not a prompt filter: a filter judges text, a control plane judges actions. Many teams need those as well. They do different jobs.
Questions to ask before you choose one
- Does it decide before the action runs, or report afterwards?
- Can I require human approval by type of action, and are there types that can never be made automatic?
- Can different agents have different levels of autonomy?
- Does it work with the agents I already use, without changing their code?
- Who holds the credentials for the tools?
- Can I check the record myself, and export it?
- What happens to a request when nobody answers?
- What does the vendor say the product does not do?
Where Operon OS Agenaut fits
Agenaut is a control plane for the agents one person or a small team runs. Claude, ChatGPT and other agents that speak MCP connect to one address. Each agent gets a level: Observer, Assistant or Trusted. Every tool has a risk class, and the class decides. Reads and ordinary writes run. Sharing outside, sending to someone new, deleting, changing settings and new forwarding rules wait for you, whatever the level. Sending, paying and deleting can never be made automatic. Every decision gets a signed receipt in the ledger.
What it does not do: it governs the tool calls that go through it, not model calls, and it does not move money. Autonomy is set per agent and by rules, not per workflow. It is in beta and holds no compliance certification. Operon OS, the enterprise platform at operonos.com, is the larger system it comes from.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- Autonomy levels for AI agents: deciding what runs without you
- Human approval by action type: read, write, share, send, delete
- Human in the loop for AI agents: approve risky actions before they run
Terms used here are explained in the glossary.