Introducing Claude Managed Agents

1. What Claude Managed Agents is — and why it exists

Building an agent on top of a raw language model API sounds straightforward until you try to do it at production scale. You add tool calling, then you need to persist memory across sessions, then you need to spin up multiple agents and have them coordinate, then you need somewhere to store intermediate state, then you need to handle retries and errors gracefully. Before long you have built a substantial platform just to support your actual product. Claude Managed Agents is that platform, built by Anthropic so you do not have to build it yourself.

The conceptual jump from the Claude API to Managed Agents is the jump from model-as-library to model-as-runtime. When you call the raw /messages endpoint you are sending a stateless request and getting a stateless response. The API knows nothing about the previous turn unless you re-send the entire conversation history yourself. Managed Agents, by contrast, maintains sessions, threads, memory, and all the orchestration logic across the full lifetime of an agent task. You define what the agent should do and what it has access to; Managed Agents handles the rest.

This distinction matters most for teams that are not trying to build developer tooling or research prototypes. If your product is an agent — if you are shipping something to end users that autonomously does meaningful work on their behalf — you need this layer. Rebuilding it from scratch takes months and introduces reliability risks in every corner you cut.

2. The four pillars of Managed Agents

Native MCP (Model Context Protocol) is built into Managed Agents at the infrastructure level. MCP is Anthropic’s open standard for connecting agents to external systems — think of it as a universal adapter between an agent and any data source or tool. With raw API usage you have to implement MCP client logic yourself. Managed Agents ships with this integrated, so connecting your agent to a filesystem, database, or external API is a configuration step, not an engineering project.

Memory is the capability that most distinguishes agents from one-shot completions. An agent that cannot remember what it did yesterday is an agent that cannot complete multi-day tasks, cannot learn user preferences, and cannot resume interrupted work. Managed Agents provides memory stores — both short-term working memory within a session and longer-term persistent memory across sessions — as a managed service. You do not need to design the storage schema, manage embeddings, or handle cache invalidation.

Tool integrations come pre-built. Web search, code execution, file reading, browser interaction — these are the tools most agent products need, and they require significant sandboxing and safety work to run securely. Managed Agents handles that hosting and security so you can give your agent tools without building the execution environment yourself.

Orchestration means multi-agent coordination is a first-class feature, not an afterthought. Real production agent tasks are often too large or too parallelizable for a single agent to handle well. Managed Agents lets you define agent hierarchies — an orchestrator that delegates to specialized subagents — and handles the communication, state sharing, and result aggregation between them.

3. Who it is for and when to reach for it

The target audience for Managed Agents is product teams at companies that want to ship agent-powered features without building the infrastructure from scratch. Think of a company like Notion wanting to embed a Claude delegate inside their workspace, or a developer tools company wanting to ship an automated code review product. These teams know what they want to build; they do not want to spend engineering cycles reinventing session management and memory stores.

There are three natural options when building with Claude, and each has its place. The raw Claude API is best for stateless tasks — classify this text, generate this draft, answer this question — where you do not need persistence or multi-step autonomy. Claude Code is best for developer-facing agentic work, especially coding tasks where a CLI or VS Code integration makes sense. Claude Managed Agents is best when you are building a production agent product that will be used by people who are not developers, or where the agent needs to do long-running work across multiple sessions with persistent memory and tool access.

🧠 Check your understanding

4 questions · your answers are saved in this browser only

1. 1. What is the primary problem Claude Managed Agents solves compared to using the raw Claude API?

   It makes Claude's responses faster and cheaper It provides a platform layer with session management, memory, tools, and orchestration so teams don't have to build that infrastructure themselves It allows Claude to be used without any API keys It only supports single-agent workflows, making them easier to reason about

   Managed Agents is a platform layer on top of the raw API that provides all the infrastructure needed for production agents — sessions, memory, tools, and multi-agent orchestration — eliminating the need for teams to build that themselves.

2. 2. Which of the following is NOT one of the four core capabilities provided out-of-the-box by Claude Managed Agents?

   Native MCP (Model Context Protocol) support Memory that persists across sessions A visual no-code agent builder interface Multi-agent orchestration

   The four core capabilities are native MCP, memory, tool integrations, and orchestration. A visual no-code builder is not part of the Managed Agents offering — it is an API-first platform.

3. 3. When should you choose the raw Claude API over Claude Managed Agents?

   When you need multi-session memory When you are building stateless tasks like text classification or one-shot generation When you need multi-agent coordination When you want native MCP tool support

   The raw API is best for stateless, one-shot tasks where you don't need persistence or multi-step autonomy. Managed Agents adds overhead that is only worthwhile when you need its infrastructure features.

4. 4. What does 'orchestration' mean in the context of Claude Managed Agents?

   Playing background music during agent tasks Scheduling agents to run at specific times of day Coordinating multiple agents in hierarchies where an orchestrator delegates to specialized subagents Converting agent outputs into structured JSON automatically

   Orchestration in Managed Agents refers to multi-agent coordination — defining agent hierarchies where a top-level orchestrator delegates subtasks to specialized subagents, with the platform handling communication and state sharing between them.

Reset quiz

🛠️ Build it yourself

Follow these exact steps to reproduce it yourself

1. Sign up for access — Managed Agents launched in private alpha in April 2026. Check console.anthropic.com for current access and waitlist options.
2. Define your agent’s scope — Write down the task your agent will complete, the tools it needs (web search? file access? code execution?), and whether it needs multi-session memory.
3. Identify your orchestration pattern — Does one agent handle everything, or do you need an orchestrator that delegates to specialized subagents? Sketch the agent hierarchy before writing code.
4. Configure MCP connections — List the external systems your agent needs to talk to (databases, APIs, file systems). Each connection becomes an MCP server configuration.
5. Define memory requirements — Decide what information needs to persist across sessions (user preferences, task history, domain knowledge) versus what is ephemeral working memory within a session.
6. Build a minimal version first — Start with a single agent, one or two tools, and no persistent memory. Get that working end-to-end before adding orchestration and memory layers.
7. Add monitoring — Production agents fail in unexpected ways. Instrument your agent sessions with logging from the start so you can debug failures and improve task success rates.