Agent control plane
What is an agent control plane?
An agent control plane is the layer that decides what AI agents may do and records what they actually did. The agents, models and tools do the work. The control plane assigns it, limits it, gates it and checks it.
Most articles on this describe a product category. This one describes a control plane that exists: the one running the AI agents that operate my companies, around the clock, across seven model harnesses. It has nine parts. You need five of them to start.
How a task moves through it
- 01RequestA task with one owner and a named outcome
- 02QueueLocked writes, claimed by one runner
- 03RunnerPicks a healthy model harness, caps the time
- 04GateIrreversible actions wait for a person
- 05ActionThe agent does the work (data plane)
- 06VerifierReads the result back from the real system
- 07LedgerOne append-only line per settlement
- 08BoardDerived live from the ledger
Seven of the eight steps are control plane. Only the action itself is the agent. That ratio is the point: the model is the easy part.
The nine parts of the one I run
Every task names its owner, the observable outcome, what the agent may and may not do, and the proof that will settle it. Written before the work starts.
One shared task file. Writers merge under a lock, never blind-write, and never strip another task's status. A claim older than six hours counts as abandoned.
Headless runners on two machines claim work every two hours, staggered so they don't collide. They run the agent, settle the task, and report.
Seven model harnesses. A harness choice is a preference: the runner rotates to the least-recently-used healthy one, with cooldowns. Expired auth or a usage limit releases the task back to the queue instead of failing it.
A 30-minute cap per task. One pause file that stops every runner at once.
Sending, publishing, spending, provisioning and migrating wait for me. A held task is counted as held, never as done. Testing and routing never remove a hold.
The runner, not the agent, reads the result back: the page contains the expected text, the object exists in the payment processor. A zero exit code proves nothing. A cited proof file that doesn't exist downgrades the task.
Every settlement appends one line to a log nobody edits. That log is the source of truth. The board people look at is a projection derived from it, never a set of cards someone pushed.
Watchers on every loop queue their own repair tasks when something breaks. Since November 2, token usage is logged per attempt, so cost per verified task can become a number instead of a guess.
What it caught, and what it didn't
Between October 1 and November 6 the control plane settled 1,436 tasks. It held 223 at the approval gate instead of shipping them. Agents marked 585 tasks done, and only one of those carried enforced read-back proof. Token cost was logged on 8% of attempts. A control plane that holds risky work but doesn't verify finished work is half built, which is why parts 7 and 9 matter most. The full ledger is here.
The minimum viable control plane
- One queue, one owner per task, writes under a lock.
- A time cap per task and one kill switch for all of them.
- A person in front of every action that cannot be undone.
- A verifier that is not the agent, reading the real after-state.
- An append-only outcome log, with your dashboard derived from it.
The rules each part enforces are written up as the AI agent governance rules, and the bar the verifier holds agents to is the definition of done for AI agents. The free checker on that page scores your agents' logs and gives you a proof rate.
Control plane vs data plane
| Control plane | Data plane |
|---|---|
| Queue, owners, task contracts | The agent working the task |
| Approval gate, limits, kill switch | Model calls and tool calls |
| Verifier, outcome ledger, board | The email, page, call or charge |
Questions
What is an agent control plane?
The layer that decides what AI agents may do and records what they actually did: the task queue, the approval gates, the limits and kill switch, the independent verifier and the outcome log. The agents, models and tools that do the work are the data plane. The control plane governs them.
What is the control plane in agentic AI?
The same idea borrowed from networking and Kubernetes: a control plane sets policy and tracks state, while the data plane does the work. In agentic AI the control plane assigns tasks, enforces approvals and limits, and verifies outcomes. The agents act inside it.
What is an AI control plane?
Usually the same thing as an agent control plane: a central place to register AI agents, set what each may do, route approvals, and log and verify what happened. Some vendors use it more broadly for all AI usage in a company, including chat tools and model access.
How is an agent control plane different from orchestration?
Orchestration decides who does the next step. A control plane decides what is allowed to happen and proves what did. You can orchestrate agents with no control plane, and that is how agents end up posting, emailing or retrying without anyone deciding.
Should you build or buy an agent control plane?
Start with the five-part minimum in your own runner, because it is a few hundred lines and you learn where your agents fail. Buy when you have many teams, compliance requirements, or agents from several vendors that need one registry. No platform replaces a definition of done.
If you want a control plane like this running inside your company, get in touch.
Get the next scorecard.
Real numbers from agents running a real company, including what broke. One email when there's something worth reading.