Claude Managed Agents Overview

4 days ago

Anthropic provides two primary methods to build with Claude: the Messages API for direct model prompting and custom agent loops, and Claude Managed Agents, a pre-built, configurable agent harness for long-running, asynchronous tasks in a managed infrastructure.
Claude Managed Agents offers a fully managed environment where Claude can autonomously perform tasks using tools like reading files, executing code, browsing the web, and running commands, with built-in performance optimizations such as prompt caching and compaction.
The Managed Agents framework is structured around four core concepts: Agent (model, system prompt, tools), Environment (configured container template), Session (running agent instance), and Events (messages exchanged between the application and agent).
To use Managed Agents, users create an agent with specific configurations, set up an environment with necessary packages and access rules, start a session, and interact by sending events to steer or interrupt the agent, with responses streamed via server-sent events (SSE).
Claude Managed Agents is currently in beta, requiring a specific header for API endpoints, and includes rate limits for create and read operations, along with organization-level spend limits and optional Claude branding guidelines for integration partners.

Hasty Briefsbeta