Agentic AI is moving from answering questions to performing actions such as installing packages, editing files, and invoking external services. Docker’s response is a three‑layer governance model—sandbox policies, MCP tool governance, and audit + visibility—that puts enforceable controls around the runtime, the tools an agent can call, and the evidence of what happened. Practitioners need to understand these layers to keep agents useful without exposing the broader environment.
Sandbox Policies: Defining Where an Agent Can Operate
Each agent runs inside a microVM that hosts its own Docker Engine, filesystem, and network stack. Docker AI Governance adds centralized policies that restrict network destinations and filesystem mounts. Network rules can whitelist domains, IP ranges, and ports, while filesystem rules dictate which host paths are visible and whether they are read‑only or read‑write. Because enforcement occurs at the proxy and mount layers, the agent cannot override the limits on its own.
Operators can query the effective policy before launching an agent:
$ sbx policy check network api.anthropic.com
Allowed: api.anthropic.com
$ sbx policy check network blocked.example.com
Denied: blocked.example.com
The output reflects the active policy, not any instruction the agent might receive, ensuring a clear answer to the question “where can the agent operate?”.
MCP Tool Governance: Controlling What an Agent Can Do
Docker’s Model Context Protocol (MCP) provides a uniform interface for agents to call external tools. Governance is applied at the MCP Gateway, where administrators register allowed MCP servers and define per‑agent capabilities using Cedar policies. A typical policy might permit read‑only tool invocations on a specific server:
permit (principal, action == MCP::Action::"invokeTool", resource) when {
resource in MCP::Server::"example" && resource.readOnly == true
};
These policies are evaluated for every request; any call lacking a matching rule is denied. If an agent bypasses the gateway and connects directly to an MCP server, the sandbox’s network policy becomes the enforcement point instead. This separation lets teams grant tool access selectively while keeping other servers or privileged actions unavailable.
Audit + Visibility: Recording What Happened
Policy enforcement alone is insufficient for security operations. Docker AI Governance emits structured audit records for sandbox decisions and session events. Each record captures the principal, action, target, and the allow/deny outcome. By aggregating these logs, security engineers gain a traceable view of agent activity, can verify compliance with policy, and can investigate incidents without relying on the agent’s internal logs.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Adopting Docker AI Governance requires updating your CI/CD pipelines to launch agents inside sandboxes, defining network and filesystem policies that match team needs, and authoring Cedar rules for MCP tool access. Teams should integrate the audit stream into existing SIEM or log‑analysis platforms to maintain visibility. Evaluate the current agent workloads to map required resources, then iteratively tighten policies while monitoring audit logs for unexpected denials. The result is a reproducible, auditable boundary that lets agents act where they are needed without exposing the broader platform.


