On this page
Control plane versus observability
Observability tells you what happened: traces, runs, tokens, costs and errors after the fact. A control plane also changes what happens next. Monitoring shows an agent in a failure loop; a control plane pauses it. Most teams start with observability and reach for control the first time an agent misbehaves in front of customers.
| Question | Observability | Control plane |
|---|---|---|
| What did the agent do? | Yes | Yes |
| What did it cost? | Yes | Yes |
| Stop or pause it now | No | Yes |
| Switch its model without a deploy | No | Yes |
| Push an approved prompt | Some tools serve versioned prompts your code fetches | Yes, with approval |
| Act automatically on spend or errors | Alerts only | Alerts and automatic pause |
The parts of an agent control plane
- Inventory: every app and agent, with owner, model and status.
- Telemetry: runs, tokens, cost, errors and heartbeats reported by each agent.
- Commands: pause, resume, stop, restart, model and prompt updates, delivered securely to the running app.
- Policy: scopes that limit which commands each app accepts, and approvals for risky changes.
- Automation: rules that notify or pause on thresholds such as spend or error rate.
- Audit: an append-only record of who changed what and when.
Where it sits: beside the model call, or in the middle
A gateway or proxy sits in the model call path: every request goes through it, which lets it block traffic but also makes it a new dependency and usually means handing it your provider keys. A control plane like Agent Control Panel (ACP) sits beside the call: the agent calls its provider directly, a small SDK reports telemetry and receives signed commands, and if the control plane is unreachable the agent keeps working. The trade-off is that enforcement happens in your process (the SDK refuses the next call), not in a network hop.
When you need one
- An agent talks to customers or takes actions with real consequences.
- More than one agent or app is in production, and nobody has the full picture.
- Model spend matters, or a provider outage would hurt.
- Someone other than the original developer needs to be able to stop an agent.
Agent Control Panel is a control plane for agents you already shipped: connect an app with the Node or Python SDK, then monitor and operate every agent from one dashboard. See also how to add a kill switch and how a control panel relates to observability, gateways and orchestration frameworks.
Frequently asked questions
What is an agent control plane?
The system that lets a team see and change what its AI agents do while they run: inventory, telemetry, commands such as pause or model swap, permissions, approvals, automation and an audit trail.
How is it different from LLM observability?
Observability shows what happened. A control plane also changes what happens next, for example pausing an agent, switching its model or pushing an approved prompt.
Does a control plane have to proxy my model calls?
No. ACP sits beside the call: your agent calls its provider directly and a small SDK reports telemetry and receives signed commands.
Which languages does ACP support?
Node.js and Python 3.10+ through official SDKs (npm and PyPI, Apache-2.0), plus a plain HTTP API.