Control plane
The place where work is defined, reviewed, coordinated, and governed across humans and agents.
The control plane is where coordination happens. In ClosedLoop.ai, the browser-based web app is the control plane, the desktop client is the execution surface, and the Claude Code plugins supply the orchestration.
Responsibilities
- Define work — create PRDs, implementation plans, and feature artifacts.
- Structure artifacts — attach context, acceptance criteria, and ownership.
- Launch execution — route commands to compute targets over the cloud relay.
- Review outputs — surface diffs, judge reports, and activity for human approval.
- Preserve decision history — maintain the audit trail of who approved what, when.
- Govern access — manage team membership, repo connections, and compute target registration.
Why ClosedLoop.ai centers this concept
Without a control plane, agent output floats free from the process that decides what should be built.
That increases local speed while increasing organizational ambiguity. Work gets done faster but scope drift, duplicated effort, inconsistent review standards, and environmental fragility all get worse.
The control plane makes the following things observable and governable:
- Which artifact launched which run.
- Which compute target executed it.
- Who approved high-risk operations.
- What the judges scored.
- Whether the PR that resulted from it was merged, and when.
How it talks to execution
The control plane does not touch your machine directly. It sends structured desktop.command envelopes over a Socket.IO relay to a specific compute target. The desktop client dispatches each command into its localhost HTTP gateway, which enforces sandbox policy, runs the operation, and streams results back. See Desktop gateway and Cloud relay.
What control means
A working control plane lets you treat agentic execution like you treat any other governed system: with audit trails, approvals, and rollback. That is the difference between a demo and a delivery model.