Skip to content
Agentic Control Plane

Agentic Data Plane vs Agentic Control Plane

David Crowe David Crowe · · 4 min read
architecture mcp
Share X HN LinkedIn

The data plane does the work. The control plane decides whether the work is allowed and keeps the record. If you’ve done any networking, you know the split. Agent systems are growing the same one, and most teams have built only the first half.

Want the full definitions? What is an Agentic Data Plane? and What is an Agentic Control Plane? each cover one layer in depth. This post sets them side by side.

What is an agentic data plane?

The data plane is everything that executes. It sends prompts to models, gets completions back, calls tools, reads databases, writes to APIs and returns results.

If an agent creates a Jira ticket, the data plane is the HTTP request to Jira. If it queries a CRM, the data plane is the query and the rows that come back. If it sends a prompt to Claude instead of GPT-4, the model call is the data plane.

It includes:

  • Model routing: picking a provider, handling failover, trading cost against latency
  • Tool execution: calling APIs, reading files, querying databases for the user
  • Context management: conversation history, retrieval, memory
  • Streaming: sending model output back to the client
  • Orchestration: multi-step workflows and sub-agents

LLM gateways (Portkey, LiteLLM, OpenRouter), agent frameworks (LangChain, CrewAI, AutoGen) and MCP servers that expose tools all live here. They decide what to do and how. They don’t decide whether it should be allowed.

What is an agentic control plane?

The agentic control plane doesn’t execute anything. It checks each action, says allow, ask or deny, and records what happened.

For a tool call, it answers questions like these:

  1. Who is asking? The call is tied to a real user or a workspace key.
  2. Is it allowed? Policy returns allow, ask (a person approves) or deny. A new workspace starts in audit mode: calls are recorded and nothing is held, apart from a short list of dangerous commands. You turn on stricter rules when you’re ready.
  3. Is the result safe to hand back? Tool output is scanned. On harnesses whose hook can rewrite results (Claude Code is one), sensitive values are redacted before they return to the model. Elsewhere the finding goes in the log.
  4. What did it cost? Model calls are priced and logged against the session.
  5. What happened? One log, with the person and the agent named on every entry.

A control plane doesn’t care which model you use or which framework runs your agents. ACP sits where the agent’s tool calls happen, as a hook in the coding agent. It checks the call whatever sits underneath.

One caveat. A hook runs on the machine where the agent runs. If someone turns hooks off in the harness, the checks stop. Plan for that, and don’t treat any client-side layer as unbypassable.

Why the split matters

In networking nobody argues for merging the switch and the controller. Merged, they’re harder to secure, audit and scale. Agent stacks repeat the mistake.

Rules scattered through application code

Without a control plane, every team does it differently. One service checks permissions in middleware. Another validates tokens in the route handler. A third relies on the API gateway for auth and loses the user’s identity on the way to the model. A fourth checks nothing, because it was a hackathon prototype that shipped.

You end up with uneven enforcement, holes in the logs, and no single place to ask who accessed what through which agent.

Data plane tools don’t solve control plane problems

LLM gateways pick models and manage cost. They don’t check that the person behind a prompt may use the tool the model wants. Agent frameworks run workflows. They don’t hold a force-push until someone approves it. That isn’t a flaw in them. A load balancer isn’t a firewall, and an LLM router isn’t an identity provider.

Auditors like separation

Frameworks such as SOC 2, HIPAA and GDPR ask you to show access controls and logs. It’s easier to show them from one dedicated layer with its own logs and policy than from if statements spread across services. Compliance-ready governance goes into what evidence looks like. This isn’t a claim that using a control plane makes you compliant.

How they work together

  1. A user sends a prompt through Claude Code, Cursor, or a custom agent.
  2. The model decides it needs a tool.
  3. The data plane routes the call toward the tool.
  4. The control plane checks it against identity and policy.
  5. If allowed, the call runs. If the rule says ask, it waits for an approval.
  6. The data plane returns the result to the model.
  7. The control plane logs the whole exchange with attribution.

The data plane handles 1 to 3 and 6. The control plane handles 4, 5 and 7.

User → LLM → Data Plane → Control Plane → Backend
                                ↓
                          Identity ✓
                          Policy   ✓
                          Redact   ✓  (where the harness allows)
                          Cost     ✓
                          Audit    ✓

You can swap your LLM gateway without touching policy. You can change frameworks without rewriting rules. A new model provider is covered because the check sits on the tool call, not inside the execution layer.

Where to start

List which parts of your stack are data plane and which are control plane. Most teams find a lot of the first (routing, orchestration, tool execution) and almost none of the second: shared API keys, no per-user logs, no one reviewing results.

You don’t need to rebuild the data plane. Add the missing layer next to it. An Agentic Control Plane gives you identity, policy, cost and an audit log without changing the tools you already use.

Get started → · Reference architecture → · Pricing →

Running coding agents? ACP records each tool call its hook sees and makes the dangerous shell commands (a force-push, destructive SQL, curl | sh) wait for you. One command, free up to 5 agents:

macOS · Linux · WSL
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
Windows · PowerShell
irm https://agenticcontrolplane.com/install.ps1 | iex

What the installer writes  ·  Getting started

Share X HN LinkedIn
Get the next data drop
What agents actually cost, new tool-surface captures, and the occasional incident post-mortem — sent when we publish something worth your inbox, not on a schedule. Unsubscribe anytime.
Share: Twitter LinkedIn
Related posts

← back to blog