Guide · definitions
What is an AI agent control plane?
Agents stopped being chatbots the day they got deploy keys and payment scopes. The control plane is the layer that makes that survivable: policy, approvals, evidence, and audit between the model and the consequences.
The definition
Borrowed from networking, where the control plane decides routes and the data plane moves packets: in agent systems, your agents and their tools are the data plane — they do the work. The control plane decides what work is allowed to have consequences. Four components, whatever the vendor calls them:
Policy — an explicit ladder from ALLOW through ALLOW_WITH_CONSTRAINTS and APPROVAL_REQUIRED up to DUAL_APPROVAL and DENY, keyed on the action's blast radius and enforced outside the model. Approvals — the surface where a human reviews the exact action, target, impact and evidence, with authorizations that expire. Evidence — checksummed artifacts (diffs, test runs, logs, metric trends) that justify a decision and become immutable once consumed. Audit — an append-only, actor-attributed record of everything the other three did.
What it is not
Not observability. Traces and dashboards describe; they do not refuse. Not orchestration. LangGraph, CrewAI and friends coordinate agents; the control plane governs their consequences, and lives outside them on purpose — guardrails inside the same code the agent runs are guardrails the agent can reach. Not a system prompt. "Never deploy without asking" is a request. A policy engine that physically routes deploys through an approval is a rule.
The state machine underneath
Concretely, a control plane shows up as extra states in the task lifecycle. VIVUU's terminal models it like this:
The waiting states are the control plane made visible: the agent has done everything policy allows and is holding at the gate with its evidence. Execution after the yes ends in verification, not celebration — and a failed verify rolls back and lands in the audit trail.
Why the interface matters as much as the engine
A control plane whose approvals are unreadable gets rubber-stamped, and rubber-stamping is autonomy without the honesty. That is VIVUU's actual bet: the decisive surface is an attention queue of decision-ready cards — what changed, why it matters, what was already done, the one decision owed — with press-and-hold approval and evidence one tap deep. The engine keeps you safe; the interface keeps you honest.
FAQ
- What is an AI agent control plane, in one sentence?
- It is the layer that sits between your agents and the systems they touch, deciding — by explicit policy rather than model judgment — what runs freely, what runs with constraints, what waits for human approval, and what is refused, while recording evidence and an audit trail for all of it.
- How is a control plane different from agent observability?
- Observability watches; a control plane decides. Tracing, token counts and latency dashboards tell you what agents did and what it cost. A control plane governs what they may do next: policy tiers, approval gates, expiring authorizations, dual sign-off. You need observability inside a control plane, but a dashboard with no enforcement is a window, not a plane.
- Is an orchestration framework (LangGraph, CrewAI) a control plane?
- No — orchestration coordinates the work (which agent runs, in what order, with what state). A control plane governs consequences, and it deliberately lives outside the orchestration code so a prompt injection or a buggy graph cannot rewrite its own guardrails. The two compose: your graph pauses at an interrupt; the control plane owns what happens during that pause.
- Do I need a control plane for a single agent?
- The moment an agent can touch something you cannot easily undo — production, customers' inboxes, money — you need at least the minimal plane: a policy ladder enforced outside the model, an approval surface with evidence, and an append-only audit trail. One agent with deploy access has the same blast radius as ten.
- Can you trust AI agents without one?
- You can trust them with reversible work. Trust for consequential work is not a feeling about the model — it is a property of the system around it: whether wrong actions are structurally hard (policy), whether humans see evidence before saying yes (approvals), and whether every act is attributable afterwards (audit). The control plane is where those properties live.
see the whole flow, live
Everything on this page — the queue, the evidence, the hold-to-approve, the audit trail — is working in the VIVUU terminal as an interactive preview on realistic mock data. No sign-up.