Agent control plane
Agent governance and operationsAI control plane · enterprise agent control plane · agent governance layer
The single layer where a company registers, authorises, watches and stops its AI agents. It doesn't do the work: it decides which agent may act, on which systems, with which permissions, and leaves an auditable trail of every session. If you can't switch an agent off from one place, you don't have a control plane. You have loose agents.
An agent control plane answers three operational questions: who each agent is, what it can touch, and what it did. In practice that's four pieces: agent identity and ownership, a permissions and tooling policy (including the allowlist of approved MCP servers), session telemetry with an auditable record, and a stop button. The term hardened in 2026 as platforms shipped. GitHub made its enterprise agent control plane generally available in February 2026: a dedicated AI administrator role, audit log entries carrying an actor_is_agent identifier alongside user and user_id, an agent_session.task event marking session start, finish and failure, and a centralised enterprise MCP allowlist (that last piece still in preview). Implementa uses the term operationally, not as a product category: the control plane is a governance function, not a licence. Buying one is not the same as having one.
How it differs from
- Agent orchestrator
- The orchestrator hands work out to agents; the control plane decides whether that agent is allowed to work at all.
- AI observability
- Observability watches and measures; the control plane also authorises, constrains and cuts.
- LLM gateway
- A gateway governs model calls; a control plane governs agents, tools and sessions.
Examples
- Cutting a support agent's access to the billing API without touching the agent's code
- Answering «who executed this change?» in an audit by naming the human the agent acted on behalf of
- Publishing the list of MCP servers your agents are allowed to call
FAQ
- Do I need a control plane with only three agents?
- You need it the moment an agent can write to a real system. With three agents a registry plus a permissions policy is enough. Building it now is cheap; retrofitting it isn't.
- Is it something you buy?
- Products implement it, but it's a governance function. Your existing platform may already cover it if it gives you agent identity, tool policy and an auditable log.