Claude Managed Agents, Anthropic’s pre-built agent infrastructure platform, entered public beta on April 1, 2026, giving developers a managed environment to run artificial intelligence tasks without building their own agent loops or execution layers.
The platform sits alongside Anthropic’s existing Messages API, which offers direct model access for custom agent development. Claude Managed Agents, by contrast, handles the infrastructure automatically, letting Claude read files, run shell commands, browse the web, and execute code inside secure cloud containers.
What Claude Managed Agents Offers
The platform is built around four core concepts: the agent itself, which defines the model, system prompt, tools, and MCP servers; the environment, a configured container with pre-installed packages and network rules; the session, a running agent instance performing a specific task; and events, the messages exchanged between the application and the agent. Built-in prompt caching and compaction features aim to improve output quality and reduce processing overhead.
Supported Tools and Capabilities
Developers using Claude Managed Agents get access to a set of built-in tools. These include Bash for running shell commands, file operations covering read, write, edit, glob, and grep functions, web search and URL content retrieval, and support for MCP servers to connect external tool providers. The platform targets workloads that run for minutes or hours with multiple tool calls, making it suited for software development automation and data processing pipelines.
How Sessions Work
To start using the platform, developers first create an agent by defining its model and configuration, then set up a cloud container environment with the required packages, such as Python, Node.js, or Go. A session is then launched referencing both configurations. Claude streams results back via server-sent events (SSE), and the full event history is stored server-side for later retrieval. Developers can also send additional events mid-session to steer or interrupt the agent’s execution.
Stateful sessions persist both the file system and conversation history across multiple interactions, removing the need for developers to manage state externally. This design targets use cases where context must carry over between steps in a long task.
Beta Access and Rate Limits
All Claude Managed Agents endpoints require the managed-agents-2026-04-01 beta header, though the official SDK sets this automatically. Access is enabled by default for all API accounts holding a valid Claude API key. Rate limits apply at the organization level: create endpoints are capped at 60 requests per minute, while read endpoints allow up to 600 requests per minute. Organization-level spend limits and tier-based rate limits also apply on top of these figures.
Three features, specifically outcomes, multiagent coordination, and memory, remain in research preview and require separate access requests. Anthropic said behaviors may be refined between releases as the beta progresses.
Branding Rules for Partners
Anthropic published branding guidelines for partners integrating Claude Managed Agents into their products. Permitted references include “Claude Agent,” “Claude” within an agents menu, and “{YourAgentName} Powered by Claude.” Partners may not use the names “Claude Code,” “Claude Cowork,” or any visual elements that mimic those products. Anthropic said products built on the platform should maintain their own distinct branding and directed compliance questions to its sales team.
