# Claude Managed Agents explained: how it works, cost and self-hosting

> Anthropic's hosted agent harness costs token rates plus $0.08 per running session-hour, and since November 12 the whole service can run on your own infrastructure.

URL: https://claudegpt.si/news/what-is-claude-managed-agents  
Published: 2026-11-15T02:55:05.170Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Tools  
Topics: Anthropic, Claude, AI agents, Agent SDK, Developer tools  
Publisher: ClaudeGPT Si

## Key takeaways

- Claude Managed Agents is Anthropic's hosted agent harness, in beta since April 8, 2026 and enabled for every Claude API account.
- It costs standard token rates plus $0.08 per session-hour, counted only while a session is running.
- Since November 12, 2026, companies can run the whole service on their own infrastructure through self-hosted Claude Managed Agents, by request through Anthropic's account team.

Claude Managed Agents is Anthropic's hosted service for running Claude as an autonomous agent. You describe the agent once (model, instructions, tools), and Anthropic runs the loop, the sandbox the agent works in and the storage for every session. It is in beta, open to every Claude API account, and costs the normal token rates plus $0.08 for each hour a session is actively running, according to [Anthropic's pricing page](https://platform.claude.com/docs/en/about-claude/pricing#claude-managed-agents-pricing).

The newest change came on November 12, 2026. Anthropic's [release notes](https://platform.claude.com/docs/en/release-notes/overview) say it launched [self-hosted Claude Managed Agents](https://platform.claude.com/docs/en/managed-agents/self-hosted): companies can now run the whole service on their own servers or in their own cloud account, not only the sandboxes. Access goes through the Anthropic account team.

## What Claude Managed Agents does

Building an agent on a plain model API means writing plumbing: call the model, run the tools it asks for, feed results back, keep and trim the history, recover from failures. The [Managed Agents overview](https://platform.claude.com/docs/en/managed-agents/overview) calls Anthropic's version of that plumbing a "harness": the agent loop, tool execution and runtime, run on Anthropic's infrastructure.

Inside a session, Claude can run shell commands, edit and search files, search the web and call MCP servers (a standard way to plug outside tools into a model). The harness also handles prompt caching and compaction, which summarizes old history so a long task fits in the context window.

Anthropic [launched it in beta on April 8, 2026](https://claude.com/blog/claude-managed-agents), promising teams would "get to production 10x faster", a figure from Anthropic and its customers, not an independent test. Notion, Rakuten, Asana and Sentry are among the named customers.

## How it works: agents, environments, sessions and events

The docs build everything on four objects:

| Object | What it is |
| --- | --- |
| Agent | The model, system prompt, tools, MCP servers and skills. Created once, reused by ID. |
| Environment | Where sessions run: an Anthropic-managed cloud sandbox, or a sandbox on your own infrastructure. |
| Session | One running instance of an agent in an environment, working on one task. |
| Events | The messages between your app and the agent: user turns, tool results, status updates. |

You create an agent and an environment, start a session that points at both, and send a user message. Claude then works on its own, calling tools and streaming progress back over server-sent events. You can steer it with more messages or interrupt it. The service keeps the full event history, so a session can resume after a pause.

Every request needs the `managed-agents-2026-04-01` beta header, which Anthropic's SDKs add for you. You can also use the REST API or the `ant` command-line tool.

Since launch, the release notes show a steady run of additions: memory (April 23), multiagent orchestration and graded outcomes (May 6), [dreams](https://claudegpt.si/news/claude-dreaming-agent-memory), an offline memory cleanup in research preview (May 6), self-hosted sandboxes (May 19), cron-scheduled sessions (June 9) and hard spending caps per session (August 7).

The service runs on all current Claude models, from Haiku 4.5 up to Opus 5.5 and the Fable and Mythos models, and is in beta on the Claude API, Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed).

## How much Claude Managed Agents costs

Anthropic bills two things, as of November 15, 2026: tokens and session runtime. There is no seat or platform fee, according to the [product page](https://claude.com/platform/claude-managed-agents).

| Charge | Rate | Notes |
| --- | --- | --- |
| Tokens | Standard model prices | Prompt caching discounts apply. US-only inference adds 10%. |
| Session runtime | $0.08 per session-hour | Counted to the millisecond, only while the session is `running`. |
| Web search | $10 per 1,000 searches | Same as the Messages API. |
| Batch discount | Not available | Sessions are interactive, so there is no batch mode. |

Idle time is free: a session waiting for your next message or a tool approval adds no runtime. The runtime charge also replaces the container-hour billing that Anthropic's code execution tool uses, so you don't pay twice for the sandbox.

Anthropic's own worked example: a one-hour coding session on Claude Opus 5.5 ($4 per million input tokens, $20 per million output tokens) that uses 50,000 input tokens and 15,000 output tokens costs $0.20 + $0.30 + $0.08 = $0.58. With 40,000 of those input tokens served from the prompt cache, it drops to about $0.43.

For most agent work, the $0.08 hourly fee is small next to the tokens.

Pricing for the self-hosted version is not published. Anthropic says to contact the account team.

## The self-hosted version, launched November 12

Until November 12, Anthropic always ran the service. The new option moves it to you:

| Option | Sandboxes | Service | Model |
| --- | --- | --- | --- |
| Cloud sandboxes (default) | Anthropic | Anthropic | Anthropic |
| Self-hosted sandboxes | You | Anthropic | Anthropic |
| Self-hosted Managed Agents | You | You | Anthropic, or a provider you choose |

The [self-hosted docs](https://platform.claude.com/docs/en/managed-agents/self-hosted) describe one deployment as a server (`cma-server`) that runs the API and the agent loop, a PostgreSQL 16 database for agents and session history, and workers that run the agent's shell and file tools inside your sandboxes. It can live in a cloud account or an on-premise data center, and it has the same API as the hosted service.

Anthropic names two reasons to choose it. You must store agents, sessions and history on your own infrastructure, or your model calls must go to Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry or your own model gateway instead of the Claude API. If you only need tools to run on your machines, the docs point you to self-hosted sandboxes instead.

It also changes data retention. The hosted service stores sessions on Anthropic's side, so it is not eligible for Zero Data Retention. The self-hosted version is the exception.

There are trade-offs. You run, secure and upgrade the server yourself. The [differences page](https://platform.claude.com/docs/en/managed-agents/self-hosted/differences) lists what the self-hosted server lacks: Anthropic's cloud sandboxes, scheduled deployments, dreams, webhooks, the Files API, and the prebuilt Word, PDF, PowerPoint and Excel skills. Nothing moves over from the hosted service, session history included.

This matters most for companies whose rules say agent transcripts, or model calls, must stay inside their own network or cloud contract. For a startup, the hosted service is still the simpler choice.

## Who Claude Managed Agents is for

The docs pitch it for long-running, asynchronous work: tasks that take minutes or hours and many tool calls, need a sandbox with network access, or run on a schedule. Anthropic's own comparison on the [Agent SDK overview](https://code.claude.com/docs/en/agent-sdk/overview) draws the lines:

| You want to | Use |
| --- | --- |
| Have Anthropic host the agent, configured through the API | Claude Managed Agents |
| Embed Claude Code's agent in your own Python or TypeScript app, in a process you run | Claude Agent SDK |
| Call the model directly and write the tool loop yourself | Messages API with a client SDK |
| Work interactively from a terminal | Claude Code |

If you want full control of the loop and your own hosting, [building on the Agent SDK or the Messages API](https://claudegpt.si/news/claude-how-to-build-an-agent) is the other route.

## What to watch

Managed Agents is still a beta, and the self-hosted server is a first version available only on request. Watch the [release notes](https://platform.claude.com/docs/en/release-notes/overview) for a general-availability date, published self-hosted pricing, and which hosted-only features reach the self-hosted server.

## Frequently asked questions

### How much does Claude Managed Agents cost?

You pay the standard token prices of the model the agent uses, plus $0.08 per session-hour while a session is running. Idle time is not billed, and web searches cost $10 per 1,000. Self-hosted pricing comes from Anthropic's account team.

### Is Claude Managed Agents free?

No. There is no seat or platform fee, but every session is billed for its tokens and its running time. Access itself is enabled by default for all Claude API accounts.

### Can Claude Managed Agents be self-hosted?

Yes. Since November 12, 2026, self-hosted Claude Managed Agents runs the service, its PostgreSQL database and the sandboxes on your own infrastructure, with model calls going to Anthropic or a provider you choose. Access is by request through your Anthropic account team.

### What is the difference between Claude Managed Agents and the Agent SDK?

Managed Agents is hosted: Anthropic runs the agent loop and stores the sessions, and you configure it through the API. The Agent SDK is a Python or TypeScript library that runs Claude Code's agent loop inside a process you operate.

## Sources

- [Claude Managed Agents overview](https://platform.claude.com/docs/en/managed-agents/overview), Anthropic
- [Self-hosted Claude Managed Agents](https://platform.claude.com/docs/en/managed-agents/self-hosted), Anthropic
- [Claude Developer Platform release notes](https://platform.claude.com/docs/en/release-notes/overview), Anthropic
- [Pricing: Claude Managed Agents pricing](https://platform.claude.com/docs/en/about-claude/pricing), Anthropic
- [Differences from the hosted service](https://platform.claude.com/docs/en/managed-agents/self-hosted/differences), Anthropic
- [Claude Managed Agents reference](https://platform.claude.com/docs/en/managed-agents/reference), Anthropic
- [Claude Managed Agents: get to production 10x faster](https://claude.com/blog/claude-managed-agents), Anthropic
- [Claude Managed Agents product page](https://claude.com/platform/claude-managed-agents), Anthropic
- [Agent SDK overview](https://code.claude.com/docs/en/agent-sdk/overview), Anthropic
