The Agent Control Plane

Infrastructure for Running AI Agents in Production

Agents can now execute actions across your business — sending emails, writing code, modifying records, triggering workflows. The infrastructure for running them exists. The infrastructure for governing what they do does not. That gap is the agent control plane.

What is an agent control plane?

An agent control plane is the governance layer that sits above AI agent execution. It does not run the agents. It governs them — defining which actions require human approval, intercepting consequential decisions before they execute, and maintaining a complete audit trail of every outcome.

The analogy is air traffic control. ATC does not fly the planes. It manages the airspace — sequencing decisions, routing traffic, issuing clearances based on a continuously updated picture of everything in flight. An agent control plane does the same for AI agent governance: it manages the decision flow above execution, not the execution itself.

Without a control plane, every agent acts independently. There is no interception point. No record of what was proposed. No mechanism to route human approvals. No pattern memory. Each agent is a plane flying without clearance.

What problems does an agent control plane solve?

Preventing irreversible actions before execution

Agents don't distinguish between reversible and consequential. Sending an email, committing code, modifying a database record — from the agent's perspective, these are equivalent outputs. A control plane applies that distinction. Consequential actions wait for approval. Routine actions run automatically.

Routing human approvals for risky decisions

Some decisions require human judgment. An agent control plane identifies those decisions before execution and routes them to the right operator — via dashboard, Slack, or any approval surface. The operator sees the full proposed action, approves or denies, and the record is written either way.

Centralizing oversight across many agents

At scale, agent operations involve multiple agents running across multiple models simultaneously. A control plane provides a single surface: one view of every agent, every action, every decision. Not per-agent logging — centralized AI agent governance across the entire fleet.

Building a structured audit trail of every action

Every gate fired, every approval, every denial is a labeled decision. What the agent proposed. What context surrounded it. What the operator chose. That accumulation becomes the decision intelligence that runs the operation — what can be automated, what requires judgment, what has never been seen before.

Where the control plane sits in the execution stack

Every AI agent execution follows the same path: a prompt goes in, a model reasons, an output comes out, an action executes. The control plane intercepts at the decision point — between output and execution.

Without the control plane, step 6 does not exist. The agent executes directly after the model responds. The operator sees the consequence, not the decision.

Decisions improve future decisions

Every action a control plane intercepts is a data point. Every approval and denial is a labeled decision — what the agent proposed, what context surrounded it, the model that produced it, the cost of the run, and what the operator chose. That data accumulates.

Over time, the pattern emerges: which actions are always approved, which are always denied, which depend on context. The control plane learns the operation. Routine decisions can be automated with confidence. Novel decisions get flagged. Risk thresholds become specific to the workflow, the agent, the operator.

This is the compounding return of running a control plane. The first gate is a checkpoint. The hundredth gate is a training signal. The thousandth gate is institutional memory.

Traditional software doesn't improve from use. An agent control plane does. Every decision routed through it makes the next decision faster, more accurate, and less dependent on human intervention. The goal is not permanent human oversight — it's the right level of oversight, continuously calibrated by the decisions that came before.

Why AI agents need a control plane

Agents are nondeterministic

Unlike traditional software, an agent does not follow a fixed execution path. The same prompt can produce different outputs, different tool calls, different actions. At scale, that variance is the risk.

Actions are often irreversible

Sending an email cannot be unsent. A committed code change has downstream effects. A deleted record may not be recoverable. An agent acting without a control plane treats all actions as equivalent. A control plane distinguishes between reversible and irreversible, routine and consequential. This is the foundation of human-in-the-loop AI.

Logs are not enough

Observability tools tell you what happened. A control plane intercepts before it happens. The distinction matters: a log of a bad decision is a record of damage. A gate before the decision is prevention. AI agent observability and AI agent governance are not the same capability.

The decision data compounds

Every gate fired is a labeled decision — what the agent proposed, what context surrounded it, what the operator chose. That accumulation becomes the decision intelligence that runs the operation. Over time, the control plane learns what requires approval and what can run automatically.

How runshift implements the agent control plane

runshift is the agent control plane for teams running AI agents in production. It is the operating system for AI agents — the layer above execution that governs what runs, what waits, and what gets routed for human judgment. It connects to any agent via AMP (Agent Message Protocol): a single signal before execution that puts runshift in the path.

Every action that passes through runshift is captured in full: the proposed action, the agent that produced it, the model that reasoned it, the cost of the run, the operator's decision, and the downstream outcome. That record is the decision model — the structured dataset of every choice made across every agent in the operation.

Gates

A gate is a human-in-the-loop approval interrupt. It fires before a consequential action executes — not after. The operator sees the full draft or proposed action, approves or denies from the dashboard or Slack, and the record is written either way. Gate policy is configurable per agent: always, on risk, or never.

relay

relay is the orchestration intelligence layer. It interprets operator intent, routes commands to agents, evaluates outputs, and surfaces what matters. When a gate fires, relay speaks first — providing context, flagging risk, suggesting edits. relay is not a chatbot. It is the operational layer between the operator and the agent fleet.

Audit trail

Every agent action, gate decision, approval, and denial is written to an immutable audit trail. Not logs — a structured record of decisions. Who approved what, when, at what cost, with what outcome. The audit trail is the foundation of the AI agent governance layer.

Frequently asked questions

What is an agent control plane in simple terms?

An agent control plane is the layer that governs AI agents before they act. It intercepts consequential actions, applies approval rules, and records every decision in an audit trail. It sits above the agent and model layers, between output and execution.

What is the difference between an agent control plane and agent monitoring?

Monitoring tells you what happened. A control plane intercepts before it happens. Agent monitoring and AI agent observability tools observe outputs and log results — they have no mechanism to prevent or approve an action before execution. A control plane sits in the execution path. The distinction is the difference between observability and governance.

How does a control plane improve over time?

Every gate fired is a labeled decision — the proposed action, the surrounding context, the operator's choice. As those decisions accumulate, patterns emerge: which actions are always approved, which require review, which have never been seen before. The control plane uses that history to route future decisions more intelligently — automating what's routine, flagging what's novel, and continuously calibrating what requires human judgment. The dataset of decisions is what makes the system smarter.

Does a control plane work with any AI model?

Yes. runshift is model-agnostic — it connects to Claude, GPT-4o, Gemini, and any custom model via AMP. The control plane sits above the model layer, not inside it. The agent signals runshift before acting, regardless of which model produced the output.

What is a gate?

A gate is a human approval interrupt that fires before a consequential agent action executes. The operator sees the proposed action — the full draft, the API call, the record modification — and approves or denies. Gate policy is configurable per agent: always (every action requires approval), on risk (relay evaluates and gates if consequential), or never (the agent runs autonomously, audit trail still written).

How does the control plane handle autonomous agents?

Autonomous agents that connect via AMP signal runshift before acting. The signal carries the proposed action, a summary, and any draft content. runshift evaluates against the agent's gate policy, fires the gate if required, and returns an approval or denial. The agent proceeds only on approval. Without AMP, the agent runs outside the control plane — runshift cannot intercept.

What is the Agent Message Protocol (AMP)?

AMP is the open protocol that connects any external agent to the runshift control plane. A single POST request before execution — carrying the agent ID, the proposed action, and optional draft content — puts runshift in the AI agent approval workflow. No SDK required. AMP works with any agent built on any framework. read amp spec →

Who is the agent control plane for?

The agent control plane is built for operators and teams running AI agents in production — founders, executives, and businesses deploying autonomous workflows at scale. Not engineers building agents. The engineers build the agents; the operators run them. runshift is the interface for the operator: plain language commands, a visual roster of every agent, and a gate that fires before anything consequential happens.

Ready to put a control plane above your agents?

runshift connects to any agent in one line. No SDK. No rebuild.