Anthropic rolls out Managed Agents

Anthropic just moved one layer up the artificial intelligence stack: instead of only selling a model, it is now selling the machinery that keeps an agent running for hours after you hit start. Claude Managed Agents entered public beta on April 8, 2026, as a hosted service on the Claude Platform. (anthropic.com) (thenewstack.io) An agent is the part that keeps taking steps after the first answer, like a travel assistant that searches flights, opens files, runs code, and comes back with a finished plan instead of one paragraph. Anthropic says most teams can already get the “brain” from a model, but still spend months building the “hands” around it. (anthropic.com) (thenewstack.io) Those “hands” are the dull but necessary parts: sandboxed code execution, checkpointing, credential management, scoped permissions, and tracing across every step. Anthropic’s pitch is that companies should hand that plumbing to Anthropic instead of building it themselves. (thenewstack.io) The company’s engineering post explains the design with an old operating-system idea: keep the interface stable even when the machinery underneath changes. Anthropic breaks an agent into three pieces — a session, a harness, and a sandbox — so it can swap the implementation of each part without breaking the product built on top. (anthropic.com) A session is the running log of everything the agent has done, like the full paper trail in a case file. A harness is the loop that decides when to call Claude and when to send Claude to a tool, while a sandbox is the sealed workspace where the model can run code and edit files. (anthropic.com) That sounds abstract until you look at what companies had been doing before this. Teams that wanted a production agent usually had to wire up secure containers, permission systems, state management, and tool execution on their own, which is why many demos never became products. (the-decoder.com) (thenewstack.io) Anthropic says Managed Agents cuts that path by a factor of 10, and outside reports describe the gap more bluntly: from months to days or weeks. The service can keep sessions alive for hours and preserve results even if the client connection drops. (the-decoder.com) (siliconangle.com) The first customers tell you who this is for. Notion is using it for delegation inside its workspace, Rakuten built enterprise agents for sales, marketing, and finance inside Slack and Microsoft Teams, and Sentry paired it with a debugging flow that writes patches and opens pull requests. (the-decoder.com) Anthropic is also charging for it like infrastructure, not like a chatbot subscription. Customers pay normal Claude token prices plus $0.08 per session-hour of active runtime, and web search adds another $10 per 1,000 searches. (the-decoder.com) (thenewstack.io) There is a catch in the fine print: for now, the service runs only on Anthropic’s own infrastructure. The public beta is available to API accounts, but companies that want the same setup through Amazon Bedrock or Google Vertex Artificial Intelligence do not have that option yet. (the-decoder.com) That makes this launch less about a smarter model and more about who gets to operate the workplace software built on top of one. Anthropic is betting that procurement teams would rather buy a managed agent service with guardrails, billing, and logs than fund another internal prototype that never survives contact with production. (anthropic.com) (thenewstack.io)