Claude Managed Agents is an AI agent service from Anthropic. It lets one AI direct other AIs to split up the work, an approach called multi-agent. Developers use it through the Claude API.

What Claude Managed Agents is

Start with the word agent. An agent is an AI that can read files, run commands, browse the web, and finish a whole chain of steps on its own. It is more than a chatbot that answers questions. For the full picture, read the complete AI Agent guide.

Claude Managed Agents is a ready-made, configurable working environment for agents, built by Anthropic and run on servers Anthropic manages. It suits long jobs that don’t need you watching. You don’t set up the environment yourself, and you don’t write code to tell the AI what to do and when. Claude can read files, run commands, and browse the web directly.

The whole service is currently in beta (a testing stage) and is enabled by default on all API accounts. If you plan to handle sensitive data that falls under regulations, such as medical data, check first that this use is allowed.

What multi-agent means

Multi-agent orchestration means one agent coordinates with other agents to finish a complex job. Each agent has its own context (the conversation the AI remembers at that moment) and can work at the same time. The docs say this can improve output quality and may also finish faster.

Here is an analogy. You own a restaurant and need to cater a banquet for a hundred guests. One person chopping, cooking, and plating would still be at it at sunrise. So you hire a head chef to run the kitchen, and the chef hands the chopping, the soup, and the dessert to different line cooks. The head chef is the main agent, and the line cooks are the other agents.

This feature is also labeled beta. You decide which of the ways to divide work are allowed, and the agent itself decides when to use them.

Three ways to divide the work

Subagents. The main agent hands tasks to helpers on its own and reads their reports. Each helper works in its own session thread (its own conversation thread), and the thread stays around after the helper finishes, so the main agent can ask follow-up questions. There is only one level of helpers. They cannot hand work down to others, and a single task can have up to 25 at the same time.

Dynamic workflows. The main agent writes a short program that calls a large group of agents in stages, has them work at the same time, and merges the results. Anthropic’s servers run the program in the background. The docs give examples such as large-scale checks, system migrations, deep research, and cross-checking data. Every agent uses tokens, so the docs recommend setting a session budget.

Advisor. Partway through a job, the main agent can ask another model for an opinion, for example to plan an approach or to do a final check. The main agent still does the hands-on work. This setting is off by default, and usage is billed at the advisor model’s rate.

When multi-agent is worth using

The docs list three fitting situations.

  • Parallel work: search several sources or analyze different files at once, then have the main agent pull it together.
  • Specialization: give each agent its own instructions and tools, for example one for security and one for documents, so you don’t pack every capability into one agent.
  • Escalation: when something is hard, ask another model for advice. That is the advisor above.

On the other hand, when a job is small, one agent doing it all costs less and uses no extra tokens. If you can’t tell whether multi-agent fits, the docs suggest reading their separate article on when to use it and when not to.

All the agents share the same working environment and files, but each keeps its own conversation history. Agents can also connect to outside tools through MCP (a standard for plugging AI into outside services; see What is MCP).

Further reading

Cover image source: Anthropic (claude.com/blog/claude-managed-agents)