Anthropic rolls managed agents public beta

Anthropic has moved one layer up the artificial intelligence stack. On April 8, 2026, it put Claude Managed Agents into public beta, which means developers can now rent not just a model, but the machinery that keeps an agent running for minutes or hours without falling over. (claude.com) A managed agent is the difference between buying an engine and buying a delivery van. Anthropic says developers no longer have to build the “agent loop,” the tool runner, the sandbox, the state store, and the permission system themselves before shipping anything to users. (claude.com) The product is aimed at jobs that do not finish in one reply. Anthropic’s documentation says Managed Agents is for long-running and asynchronous work, where Claude can read files, run shell commands, browse the web, execute code securely, and keep working inside a cloud container after the first prompt is over. (platform.claude.com) Anthropic breaks the system into four parts so developers can swap pieces without rebuilding the whole thing. An “agent” is the model plus prompt and tools, an “environment” is the container with packages and network rules, a “session” is the running job, and “events” are the messages and tool results that move through it. (platform.claude.com) That sounds dry until you look at the failure mode it is trying to remove. Anthropic says shipping a production agent usually means months of work on sandboxed code execution, checkpointing, credential management, scoped permissions, and tracing before a company even gets to the user-facing feature. (claude.com) The company is also making a bet that agent infrastructure will age faster than the models themselves. In an engineering post published the same day, Anthropic says the tricks that helped Claude Sonnet 4.5, including resets to avoid what it called “context anxiety,” became unnecessary on Claude Opus 4.5, so hard-coded harness logic can turn into dead weight as models improve. (anthropic.com) That is why Anthropic keeps talking about “stable interfaces.” Its engineers compare Managed Agents to an operating system abstraction, where developers code against durable ideas like sessions and files while Anthropic changes the harness and runtime underneath. (anthropic.com) This launch also fits a longer Anthropic push from raw models toward tools that do work on their own. In October 2024, the company released computer use in public beta so Claude could click, type, and navigate software like a person; Managed Agents turns that kind of capability into a hosted service developers can run in production. (anthropic.com) Anthropic’s newer models are being pitched for exactly this kind of workload. When it introduced Claude Sonnet 4.6 on February 17, 2026, the company said the model improved agent planning, long-context reasoning, and computer use, and described it as a fit for long-running autonomous workflows. (anthropic.com) So the real product here is not just Claude answering questions. It is Anthropic trying to become the company that hosts the worker, stores the memory, runs the tools, survives the crashes, and lets a developer treat an agent more like a cloud service than a weekend prototype. (claude.com)