Stack layer · Govern, secure, and observe every agent

The AI Agent Control Plane

The control plane sits outside the tools that build agents. It answers which agents exist, who they act as, what they may do, whether they are behaving, and which calls they are allowed to make.

44 current tools in 5 categories · see the whole stack

Categories in this layer

Suites that span several categories

Each tool is listed under its main function. These also cover two or more other categories in this layer.

How the industry frames the control plane

There is no single standard yet. These are the framings this grouping follows; gateways are included as the point where policy is enforced on every model and tool call.

  • Forrester: Defines an agent control plane that inventories, governs, orchestrates and assures agents across vendors, separate from the planes that build agents and orchestrate processes. Source
  • Microsoft: Positions Agent 365 as the control plane for agents, organised as observe, govern and secure. Source

Recent updates in this layer

  • 2026-09-18
    Amazon Bedrock AgentCore

    AWS released the new AgentCore Runtime (V2) with consumption-based memory billing and about 2-second P75 cold starts, in five regions.

    Source
  • 2026-09-17
    MLflow

    MLflow 3.16.1 removed the default basic-auth admin password (a security fix for self-hosted servers) and added scorer timeouts.

    Source
  • 2026-09-16
    NeMo Guardrails

    NeMo Guardrails 0.24.1 fixed benchmark concurrency measurement, Nemotron 3.5 response parsing, server overload handling, and combined-config loading.

    Source
  • 2026-08-27
    agentgateway

    agentgateway 1.5.0 added API-key-scoped LLM budgets, native Gemini APIs, and guardrails on tool calls, with breaking JWT and token-count changes.

    Source