AI product
Claude Managed Agents is Anthropic's managed agent infrastructure: a configurable agent harness for running Claude on long-running and asynchronous tasks. An agent combines a model, system prompt, tools, MCP servers, and skills; an environment specifies an Anthropic-managed cloud sandbox or a self-hosted sandbox; and a session runs the agent and produces persisted events. Applications send user messages as events, while Claude runs tools and streams responses through server-sent events; sessions can be steered or interrupted, and their event history can be retrieved.
The service provides managed tool execution and sandbox infrastructure, including shell commands, file operations, web search and fetch, and MCP connections. It includes prompt caching, compaction, persistent filesystems, conversation history, and server-side storage, so users do not need to build their own agent loop, sandbox, or tool-execution layer. The documentation also describes a design that separates the agent's harness and model from its sandboxes and tools, with credentials handled outside the sandbox through an egress proxy.
Claude Managed Agents is in beta and requires the managed-agents-2026-04-01 beta header. Stateful sessions are not currently eligible for Zero Data Retention or HIPAA Business Associate Agreement coverage.
4 uses taken from transcripts — each links to the moment in the video.
Anthropic's managed agent infrastructure. It lets you define an environment for a sandbox, create an agent with a model, system prompt and toolset (like the joke-teller demo), and run sessions that produce an append-only event stream. Anthropic redesigned it to decouple the 'brain' (harness plus Claude) from the 'hands' (sandboxes, MCP tools), keeping credentials out of the sandbox via an egress proxy and cutting time to first token by about 90%.
API/SDK platform for building and operating agents: sessions with custom tools, multi-agent sub-sessions, outcome definitions, memory stores, and deployments that run an agent on a schedule with attached resources
Anthropic's managed offering for running agents: sessions with memory stores, MCP channels, custom tools, sandboxed execution, outcome-based exit criteria for long-running work, and subagent/coordinator patterns. Discussed throughout the Q&A, including tenant isolation via workspaces and sharing resources across an org.
Hosted agent platform: creates agents and sandbox environments, runs sessions over an event stream, and supplies skills, subagents, custom tools, outcome grading, shared memory stores, vaults, MCP tunnels and scheduled deployments — infrastructure, hosting, scaling and event-log storage are managed for you.
4 in the library.