What Are Claude Managed Agents?

Before we get into the new features, it helps to understand what Claude Managed Agents are and how they’re structured.

Claude Managed Agents is Anthropic’s hosted infrastructure layer for running AI agents.

You focus on defining what the agent does. Anthropic’s infrastructure handles keeping it running.

The agent loop has always lived on Anthropic’s infrastructure.

Previously, when your agent called a tool — run code, accessed a file, triggered a browser action, that execution also happened in Anthropic’s environment.

Your data moved outside your perimeter to make it work.

Anthropic’s cloud → orchestration, reasoning, context, error handling

Your infrastructure → tool execution, file access, code runs, sensitive operations

Self-Hosted Sandboxes

A sandbox in traditional software development is an isolated environment for testing — a safe space where code runs without touching anything outside it.

In the context of Claude agents, it serves the same isolation role but in reverse.

With self-hosted sandboxes, you now control where that isolated environment lives.

When you configure a self-hosted sandbox, your team owns:

The runtime image — what packages, dependencies, and tools are pre-installed

Resource sizing — CPU, memory, and capacity scaled to the workload

Network policies — what the agent can and can’t reach from inside the sandbox

Audit logging — your existing security tooling applies automatically since execution is in your perimeter

Files and repositories stay inside your environment. They don’t leave.