Every agent has to ask permission. The control plane is who grants it.
Before an AI agent can do anything real — read a record, move money, close a ticket — something has to decide whether it’s allowed, what it can touch, and whose rules it follows. That layer is the control plane.
It sounds like plumbing. It’s actually where the next decade of lock-in is being decided. Here’s the part a CIO needs to hold onto.
Your agents can’t get to your systems directly. They pass through the control plane, every time. The only question that matters is who owns that gate.
Is your gate open, or is it someone else’s?
Every vendor sits somewhere on this line. The further right, the more your agents have to route through their gate, on their terms, at their meter.
Agents route on open protocols. You keep identity and the kill-switch. Most flexible, most work to run yourself.
Open on the surface, but real access is metered through the vendor's pathway. Convenient, and quietly billable.
To act on your own data, agents must route through their control plane. Easiest to buy, hardest to leave.
Most enterprise vendors are drifting rightward, privileging their own agents and metering everyone else’s. That’s not wrong, it’s strategy. You just need to know where each of yours sits before you stake an architecture on it.
“Control plane” means two different things
Two vendors can use the exact same phrase and be selling completely different things. When someone says “control plane,” ask which floor they mean.
Govern the agents doing the work
The gate above: identity, routing, governance, the kill-switch. This is what most people mean, and it’s the one this page is about.
Build the software and the agents themselves
“Services-as-Software”: firms like Factory.ai and 8090 that build the system for you. They also say “control plane,” but they mean how software gets made, not how agents get governed.
Whoever owns the control plane governs your agents. Don’t let your governance become someone else’s product.