What Is Claude Managed Agents?

Claude Managed Agents is a platform for running Claude as an autonomous agent.

Until now, building an AI agent meant implementing everything yourself: the agent loop (the repeated cycle of LLM call → tool execution → returning results), a tool execution environment, sandboxing, and state management. With Managed Agents, Anthropic manages all of that for you. All the developer has to do is “configure the agent” and “send messages.”

How It Differs from the Messages API

The Messages API is the most basic API for calling Claude: you send a prompt and get a single model response back.

	Messages API	Claude Managed Agents
Overview	Direct prompt access to the model	A configurable agent harness running on managed infrastructure
Agent loop	You implement it yourself	Provided by Anthropic
Tool execution	Runs in your own code	Runs inside a cloud sandbox
State	Stateless (send the full history every time)	Stateful (history and files are kept server-side)
Best suited for	Fine-grained control, custom loops	Long-running tasks, asynchronous processing

Reference: Claude Managed Agents overview

The Four Concepts

You can understand Managed Agents by grasping these four concepts. In today’s hands-on, we will build things in this order as well.

Concept	Description
Agent	A reusable, versioned configuration that bundles the model, system prompt, tools, MCP servers, and skills
Environment	The configuration of where sessions run: an Anthropic-managed cloud sandbox, or self-hosted on your own infrastructure
Session	An instance of an agent running inside an environment. It executes tasks and produces artifacts
Event	A message exchanged between your application and the agent (user messages, tool execution notifications, status updates, etc.)

Reference: Agents / Environments / Sessions / Events

How It Works

1. Create an agent

   Define the model, system prompt, and tools. Once created, it can be referenced any number of times by its agent ID.

2. Create an environment

   Configure where the agent runs (the sandbox configuration).

3. Start a session

   Launch a session that references the agent and environment. Each session is assigned its own isolated Linux container.

4. Send events and stream the response

   When you send a user message, Claude autonomously executes tools and streams the results back via SSE.

5. Course-correct or interrupt

   You can send additional messages to steer the agent without stopping the run. You can also interrupt and stop it.

Reference: Quickstart

Built-in Tools

Specifying agent_toolset_20260401 in the agent configuration enables the following set of tools.

Tool	Description
bash	Run shell commands inside the sandbox
read / write / edit	Read, write, and edit files
glob / grep	File search and text search
web_search / web_fetch	Web search and fetching content from URLs

You can enable or disable tools individually, and configure permission policies that require approval before execution. In addition, you can use MCP server connections, your own custom tools, and skills.

Reference: Tools

Watch Sessions in Real Time in the Console

When you create a session, a trace view becomes available in the Claude Platform console. You can see in real time which tools the agent called and with what arguments. If you know the session ID, you can open it at the following URL, so we recommend keeping the terminal and the console side by side.

https://platform.claude.com/workspaces/default/sessions/{session ID}

Reference: Session operations

Pricing

1. Tokens consumed by the model within a session are billed at the standard API rates

   Optimizations such as prompt caching and compaction are built into the harness

   Model	Input / 1M tokens	Output / 1M tokens
   Claude Opus 4.8 (claude-opus-4-8)	$5.00	$25.00
   Claude Sonnet 5 (claude-sonnet-5)	$3.00 ($2.00 until 2026-08-31)	$15.00 ($10.00 until then)
   Claude Haiku 4.5 (claude-haiku-4-5)	$1.00	$5.00

2. Session uptime ($0.08 per hour)

   Uptime is metered in milliseconds only while the agent is actually working (running); you are not billed while it is idle waiting for input.

Reference: Pricing