What I learned building a control plane for coding agents (Claude Code, Codex, Pi, Kimi) running on multiple machines

Once I ran more than one coding agent, I lost track of who was doing what, on which machine, and what they had changed. So I built a small control plane for them. A few lessons that might help if you're building something similar: The control plane should never run agents. Mine only queues work. Daemons on each machine poll for it and run the agent CLIs, either on the host or in a sandbox. That kept the trust boundary simple. Approvals must come from a person, never from anything the agent can reach. When a runtime asks for permission to run a command or edit a file, the run pauses until a signed-in human answers. "Alive" is not "listening". A daemon can keep sending heartbeats while its command poll is stuck. I track both and only dispatch to machines that are actually polling. Use one source of truth. Sessions and tasks are event-sourced and everything else is derived, which made handoffs, retries and "what happened here?" much easier. Treat agents as named teammates with a role, runtime and skills, not anonymous CLI calls. Assigning work gets much easier. It supports threads, tasks, scheduled automations, and teams such as one agent fixing a bug and another reviewing it. How are you tracking and approving what your agents do across machines? I'll put the repo in the first comment.