Anthropic Opens Claude Managed Agents to Public Beta for AI Workflows
The hosted service supports persistent sessions, sandboxing, tool permissions and execution tracing for autonomous and multi-agent workflows.

Anthropic opened Claude Managed Agents to public beta April 8, 2026, giving developers hosted infrastructure for long-running autonomous agents that can retain state, use tools and continue operating through disconnections.
The service provides composable application programming interfaces for agent state, memory, permissions and scheduled execution. Sessions can run autonomously for hours while preserving conversation history, sandbox state and outputs on the server.
Managed Agents supports both individual tasks and multi-agent workflows. A May 19, 2026 update added memory, outcomes and multi-agent orchestration to the public beta.
The platform handles sandboxing, authentication, tool execution, session state, scoped permissions and execution tracing. Developers can use cloud sandboxes managed by Anthropic or run sandboxes on their own infrastructure.
Permission policies control how agents interact with tools. A policy can automatically allow a tool call, require user approval or have the server evaluate whether the action should proceed. Under the auto policy, the server can allow, deny or pause a request for approval, while calls classified as high risk can be denied.
The ant beta:sessions connect command lets users follow sessions live, send messages, interrupt agents and approve or deny waiting tool calls. The command requires CLI version 1.32.0 or later.
In a company-run evaluation of structured file creation, successful outcomes rose by as much as 10 percentage points versus a conventional prompting loop. The result was presented as an internal performance measure rather than an independently validated benchmark.
Managed Agents is designed as infrastructure for running Claude as an autonomous agent, moving operational controls for long-running tasks into a hosted service. This includes session management, tool execution, sandboxing and permission handling.
Sessions and outputs are stored server-side. The service is not currently eligible for Zero Data Retention or HIPAA BAA coverage, which may affect organizations handling sensitive information.
Access is enabled by default for all API accounts. Every Managed Agents endpoint requires the managed-agents-2026-04-01 beta header.