Claude Managed Agents: Getting Started

About 5 minutes

Developers, AI engineers, and platform teams considering managed agent infrastructure

Claude Managed Agents is Anthropic’s managed runtime for agent execution. Instead of building your own agent loop, tool execution layer, sandbox, and session management, you use it as an API for long-running asynchronous Claude tasks. As of July 2026, it is in beta, and Managed Agents API calls require the managed-agents-2026-04-01 beta header.[1]

What Are Claude Managed Agents?

An AI agent is an AI program that carries out tasks by following instructions and using tools. For example, an agent might read files, run commands, search the web, and produce a report without requiring each step to be manually driven.

Traditionally, running this kind of agent requires infrastructure for tool execution, file state, conversation history, and recovery behavior. Claude Managed Agents lets you use an Anthropic-managed cloud sandbox or a self-hosted sandbox on infrastructure you control.[1]

graph TD
  USER["Application\n(starts a session)"] --> API[Managed Agents API]
  API --> MANAGED[Claude Managed Agents\nruntime]
  MANAGED --> CLAUDE["Claude model\n(reasoning and generation)"]
  MANAGED --> TOOLS["Built-in tools\n(Bash, files, web, MCP)"]
  MANAGED --> STREAM["Event stream\n(SSE responses)"]

  style MANAGED fill:#e8f4fd,stroke:#4a9eda

In one sentence: Managed Agents is managed infrastructure for running long-running asynchronous Claude agents.

Key Concepts

Agent, Environment, Session, and Events

Claude Managed Agents is built around four concepts.[1]

Concept	Role
Agent	Defines the model, system prompt, tools, MCP servers, and skills
Environment	Defines where sessions run: an Anthropic-managed cloud sandbox or a self-hosted sandbox
Session	A running agent instance performing a specific task
Events	Messages exchanged between the application and the agent, such as user turns, tool results, and status updates

Scheduling

Managed Agents supports scheduled deployments for recurring agent runs. Use this for weekly research, repository checks, report updates, or other work with a clear cadence.[1]

Built-In Tools

Managed Agents gives Claude access to built-in tools.[1]

Bash
File operations: read, write, edit, glob, and grep
Web search and fetch
MCP servers

Session Operations and Streaming

An application starts a session, sends events, and receives in-progress output through Server-Sent Events (SSE). It can also steer the agent with additional events or interrupt execution.[1]

Use Cases

Long-Running Research and Report Generation

Managed Agents fits tasks that combine multiple web searches, file reads, command runs, and generated outputs. For example, an agent can gather public information, inspect local files in a sandbox, and write a report file.[1]

graph LR
  START[Start session] --> READ[Fetch files and web context]
  READ --> ANALYZE[Claude analyzes]
  ANALYZE --> WRITE[Write output file]
  WRITE --> STREAM[Stream progress with SSE]

Asynchronous Inquiry Classification

sequenceDiagram
  participant APP as Application
  participant AGENT as Managed Agents
  participant CLAUDE as Claude
  participant API as Ticket API

  APP->>AGENT: Start session
  AGENT->>CLAUDE: Classify inquiry
  CLAUDE-->>AGENT: Return classification
  AGENT->>API: Send result to external system
  AGENT-->>APP: Stream status events

Your application starts a session with the inquiry text, available tools, and expected output format. Claude performs classification or summarization with tools, while the application receives progress through the event stream.[1]

Recurring Agent Work

Use scheduled deployments when work needs to run on a cadence: weekly research, repository inspection, report generation, or similar recurring jobs.[1]

Comparison

Claude Managed Agents vs Claude Agent SDK

Dimension	Claude Managed Agents	Claude Agent SDK
Form	Managed runtime from Anthropic	SDK used from Python or TypeScript
Core model	Agent, Environment, Session, Events, sandboxes	The same agent loop, built-in tools, hooks, MCP, permissions, and sessions that power Claude Code
Best for	Long-running asynchronous tasks, stateful sessions, scheduled runs	Custom agents embedded inside your application
Runtime	Anthropic-managed cloud sandbox or self-hosted sandbox	Your application runtime
Primary user	Teams that want managed agent infrastructure	Developers who want programmatic agent control

Claude Managed Agents vs Claude Code

Dimension	Claude Managed Agents	Claude Code
Runtime	Managed sandbox	Terminal, IDE, desktop, and web
Primary use	Long-running asynchronous agent execution	Interactive coding work with a developer
Control surface	API sessions and events	Claude Code conversation, tools, settings, and hooks
State	Server-side session history and sandbox state	Local working directory and Claude Code session

Getting Started

Prepare a Claude API key.
Create an Agent with the Managed Agents API and define its model, system prompt, tools, MCP servers, and skills.
Create an Environment for where the agent runs.
Start a Session using the Agent and Environment.
Send Events and stream responses with SSE.[1]

When to Choose Managed Agents

Good Fit

Long-running asynchronous tasks that use multiple tools
Work where sandboxing, tool execution, conversation history, and file state should be managed
Stateful sessions or scheduled execution
Applications that need to stream agent progress through an API

Poor Fit

A single synchronous model response is enough
You need complete control over the agent loop inside your application
Zero Data Retention or HIPAA BAA coverage is required. The official documentation states that Managed Agents is currently not eligible because it is stateful by design.[1]

Summary

Claude Managed Agents is useful when you want to run long-running asynchronous Claude agents on managed infrastructure.

It is organized around Agent, Environment, Session, and Events.
It can use Bash, file operations, web search/fetch, and MCP servers.
It can run in an Anthropic-managed cloud sandbox or a self-hosted sandbox.
It is a beta feature, so confirm the beta header and data retention conditions before using it.

Use Claude Agent SDK when you need fine-grained programmatic control inside your application. Use Claude Code when a developer wants to work interactively with a repository.

Claude Features & Product Lineup — full product overview and decision guide
Multi-Agent Design Patterns — Orchestrator-Worker, Pipeline, Swarm, and more
Claude API and Prompt Caching — direct API usage with the Anthropic SDK

This page’s Managed Agents features, limitations, and core concepts are grounded in the official documentation.[1]

References

Anthropic, Claude Managed Agents overview

For the latest releases and updates, check the official website and official documentation.

Quiz

Anthropic's Safety Philosophy and Claude Design

Claude Code Level Up — Portfolio Site in Practice