Anthropic ships managed agents

Anthropic launched Claude Managed Agents on April 8, offering a hosted way to run autonomous Claude systems without building the runtime from scratch. (claude.com) The product is in public beta on the Claude Platform, and Anthropic says it handles the production work that usually slows deployments: sandboxed code execution, checkpointing, credential management, scoped permissions, and tracing. (claude.com) In Anthropic’s setup, an agent is the model, prompt, tools, Model Context Protocol servers, and skills; an environment is the container template it runs in; and a session is the live instance doing the work. Developers create those pieces, then send events into the session and stream status and outputs back. (platform.claude.com) Anthropic is selling the service as a way to separate the model’s reasoning from the machinery around it. Its engineering team says Managed Agents virtualizes three parts of the stack: the session log, the harness that calls Claude and routes tools, and the sandbox where code runs and files are edited. (anthropic.com) That pitch lands at a moment when more companies want agents that can do work across applications, not just answer questions in a chat box. Anthropic wrote on April 9 that agents now write and execute code, manage files, and complete tasks across multiple applications, which creates both new business demand and new governance problems. (anthropic.com) Anthropic has been arguing for months that the hard part is not only model quality but the loop around the model. In a December 19, 2024 post, it said many teams were better served by simple workflows first, because agentic systems add cost and latency even when they improve flexibility. (anthropic.com) The new service also points to where Anthropic thinks enterprise demand is heading. Its docs include a research-preview feature for multiagent sessions, where one coordinating agent can call specialized agents in parallel, each with its own isolated context, while sharing the same container and filesystem. (platform.claude.com) For companies that need outside software access, Anthropic’s production cookbook describes using vaults to store per-user credentials and Model Context Protocol toolsets to connect services such as GitHub, Slack, Linear, Stripe, Notion, Salesforce, and Asana. Anthropic says those calls can run server-side from inside the sandbox instead of bouncing through the customer’s app each time. (platform.claude.com) The immediate test is whether managed infrastructure can make agents easier to ship without making them harder to control. Anthropic’s answer, at launch, is a beta API that tries to package the plumbing, the memory, and the execution environment into one service. (claude.com)