Claude Managed Agents overview
Claude Managed Agents is Anthropic's hosted agent runtime: you define an agent and a sandbox, and Anthropic runs the agent loop, executes tools, and keeps session state for your application.
Claude Managed Agents (sometimes searched as Anthropic Managed Agents) is a service on the Claude Developer Platform that runs AI agents for you. You describe the agent, and Anthropic runs the loop that calls Claude, executes tools, stores progress, and recovers from failures on its infrastructure.
Your application creates a session, sends a task, and listens to an event stream while Claude works — reading files, running commands in a sandbox, calling MCP tools, and reporting back. The work can run for minutes or hours, and you can come back to the same session later.
Anthropic's Managed Agents overview is the
canonical reference. CMA is in beta behind the
managed-agents-2026-04-01 header, and it is also offered on
Claude Platform on AWS and Claude Models on Google Cloud (Anthropic
managed). Recurring runs on a cron schedule are available through
scheduled deployments.
A versioned configuration: the Claude model, system prompt, tools, MCP servers, skills, and any agents it may delegate to. Agents hold no conversation state. Agent docs
The sandbox specification — packages, networking rules, and whether tools run in Anthropic's cloud sandbox or a self-hosted sandbox you operate. Environment docs
One running conversation between your application and an agent in an environment. Sessions are long-running: they can work for a long time, go idle, and be resumed later. You can send more user events to steer a running session or interrupt it. Session docs
The append-only stream of what happened: user messages, agent messages, tool use and results, and status changes such as running and idle. You stream it over SSE. Events docs
Both give Claude an agent loop with tools. The difference is where the loop runs and who owns everything around it.
| Claude Managed Agents | Claude Agent SDK | |
|---|---|---|
| Where the harness runs | Anthropic's infrastructure | Your process |
| Session state | Hosted, durable, resumable | Yours to persist |
| Sandbox | Anthropic cloud sandbox or self-hosted sandbox | Wherever your process runs, unless you add one |
| Scaling and recovery | Handled by Anthropic | Yours to build |
| Billing | Tokens + session runtime | Tokens + your own compute |
| Models | Claude | Claude |
Choosing between them, and when to move from one to the other, is covered in Claude Agent SDK vs managed agents.
The code that turns a model into an agent: it builds context, calls the model, dispatches tool calls, and decides when to compact or stop. In CMA, Anthropic writes and runs the harness.
Sessions outlive a single request. They move between running and idle, survive infrastructure failures, and keep their event history, so asynchronous work does not depend on one open connection.
Anthropic's engineering write-up describes separating the brain (model and harness) from the hands (sandboxes and tools) behind stable interfaces, so each can fail, scale, or be replaced independently.
Agents run Claude models. This becomes a constraint if you route tasks across vendors, need an open-weight model for some workloads, or want a fallback provider.
Anthropic operates the harness and stores session state, sandbox state, and event history server-side, also when you reach CMA through AWS or Google Cloud. Self-hosted sandboxes move tool execution to your infrastructure, and Anthropic keeps running the control plane. Because sessions persist, CMA is not eligible for Zero Data Retention or HIPAA BAA coverage, including with self-hosted sandboxes. You can read and configure the agent, but the harness code itself is closed.
You pay tokens plus session runtime. For long-running or always-on agents the runtime line grows. See Claude Managed Agents pricing.
Fast time to production, no runtime to operate, and a harness with prompt caching and compaction built in by Anthropic. For many teams those benefits outweigh the limits.
Consider something else when one of the limits above is a hard requirement: you need several model vendors, the control plane must run on your infrastructure, or you need to change how the harness works. Open managed runtimes such as OpenMA keep the same agents, environments, sessions, and events shape while letting you choose the model, harness, sandbox, and deployment.
Multi-model and self-hosted options are compared in Claude Managed Agents alternative, or go straight to Claude Managed Agents vs OpenMA.
If deployment control is the constraint, read self-hosted AI agents. If licensing and model choice are, compare open-source AI agents.
It is Anthropic's hosted service for running AI agents. You describe the agent and its sandbox. Anthropic runs the agent loop, executes tools in a sandbox, keeps the session state, and streams results back to your application.
Yes. Anthropic markets the product as Claude Managed Agents; people also search for it as Anthropic Managed Agents.
No. Claude Code is a coding agent a developer uses directly. Claude Managed Agents is an API platform your application calls to run agents of any kind, and it is billed by usage.
Standard token rates for the Claude model plus a session-runtime fee while a session is running. See the Claude Managed Agents pricing breakdown for a worked example.
No. Claude Managed Agents runs Claude models. If you need several model vendors behind the same agents-and-sessions API, look at open, model-agnostic managed runtimes.
Open-source and self-hosted options compared.
Session runtime, tokens, and a worked monthly example.
In-process harness vs hosted runtime, and when to migrate.
Side-by-side feature matrix.
