Anthropic opens Managed Agents beta

Most companies can get an artificial intelligence agent to work in a demo. The part that breaks is everything around the demo: the loop that keeps it running, the container it runs inside, the keys it uses, and the logs you need when it does something strange. (anthropic.com) (platform.claude.com) Anthropic’s answer is a new public beta called Claude Managed Agents, released on April 8, 2026, that runs those long-lived agents on Anthropic’s own platform instead of making each customer stitch the plumbing together themselves. Anthropic says the service is meant for “long-horizon” work, which means tasks that take many steps and keep state over time instead of answering in one shot. (anthropic.com) (siliconangle.com) To see what changed, it helps to separate a model from an agent. A model is the brain that writes the next token, while an agent is the whole worker around that brain that can call tools, remember prior steps, and keep going until the job is done. (platform.claude.com 1) (platform.claude.com 2) Anthropic already had a do-it-yourself route called the Claude Agent Software Development Kit, where developers could run the agent loop in their own code with built-in tools like file editing, shell commands, and web search. Managed Agents moves that loop into Anthropic’s cloud, so the customer defines the agent and environment while Anthropic runs the session. (platform.claude.com 1) (platform.claude.com 2) The service is built around three pieces. An agent is the reusable setup, an environment is the container template with packages and network rules, and a session is the live run that produces events as the task unfolds. (platform.claude.com) Those details sound dry until you hit failure. Anthropic’s engineering post says earlier agent “harnesses” baked in assumptions about what Claude could not do, and those assumptions went stale as models improved, so the company redesigned the system around interfaces it hopes will stay stable even when the internals change. (anthropic.com) Anthropic also says it split the system so the session log, the tool-calling loop, and the sandboxed execution environment can be swapped independently. That is the software version of replacing an engine without rebuilding the whole car, and it is meant to avoid fragile one-off setups that die when a single container fails. (anthropic.com) The beta already exposes concrete controls that show who this is for. The quickstart uses a required beta header named `managed-agents-2026-04-01`, creates versioned agents through a dedicated Agents Application Programming Interface, creates container templates through an Environments Application Programming Interface, and runs stateful sessions through a Sessions Application Programming Interface. (platform.claude.com 1) (platform.claude.com 2) That makes deployment faster, but it does not erase security work. Anthropic’s own docs show an environment can be created with “unrestricted” networking, which means customers still have to decide what the agent should be allowed to reach, what credentials it should hold, and how tightly those permissions should be scoped. (platform.claude.com) That is why this launch is really about where responsibility moves. Anthropic is taking over more infrastructure, but the customer still owns the policy layer: least-privilege access, audit trails, and the judgment about whether an autonomous agent should touch production systems at all. (wired.com) (anthropic.com) So the pitch is not “agents are solved.” The pitch is that Anthropic wants to sell the operating system for agents, the same way cloud providers sold the operating system for web apps, and the bet is that companies would rather rent that machinery than build it from scratch every time. (anthropic.com) (platform.claude.com)