Anthropic managed agents
Anthropic launched a hosted platform called Claude Managed Agents to help companies build and run autonomous AI agents without running the whole stack themselves. The service targets secure, production‑ready deployments with built‑in governance and early adopters named include Notion and Rakuten, signalling Anthropic’s push to capture more of the enterprise operational layer (the-decoder.com) (technobezz.com).
Anthropic is trying to sell companies the part of artificial intelligence work they usually build themselves: the machinery that keeps an agent running after the demo ends. On April 8, 2026, it put that machinery into public beta as Claude Managed Agents on the Claude Platform. (claude.com) Most companies can already make an agent answer a question in a browser tab. The hard part is everything around it: secure containers, saved state, permissions, tracing, and recovery when a tool call fails. (claude.com) Anthropic’s pitch is that customers should define the job, the tools, and the guardrails, and Anthropic should run the rest on its own infrastructure. The company says that can cut the trip from prototype to production from months to days, or about 10 times faster. (claude.com) The key idea is a hosted “harness,” which is the loop that decides when the model should search the web, run code, use a file, or try again after an error. Anthropic says that harness is built in, so customers do not have to keep rewriting their agent logic every time models improve. (anthropic.com) Anthropic’s engineers describe three building blocks under the hood: a session, a harness, and a sandbox. A session is the running log, a harness is the traffic cop, and a sandbox is the locked room where the agent can safely run code and edit files. (anthropic.com) That design comes from a problem Anthropic says it hit internally: agent scaffolding ages badly when model behavior changes. In one example, the company says a workaround it built for Claude Sonnet 4.5 became unnecessary on Claude Opus 4.5, so the old harness turned into dead weight. (anthropic.com) The product is aimed at long-running jobs, not just one-shot prompts. Anthropic says sessions can run for hours, survive a dropped connection, and use built-in tools including bash commands, file operations, web search, and connections to outside services through Model Context Protocol servers. (the-decoder.com) (claude.com) The early customer list shows the kinds of work Anthropic wants to own. Notion is using it for delegated work inside its workspace, Rakuten built agents for sales, marketing, finance, and engineering inside Slack and Microsoft Teams, and Sentry paired it with a debugging flow that can write patches and open pull requests. (the-decoder.com) Anthropic is also charging for the runtime itself, not just the model tokens. The Decoder reports a price of $0.08 per session hour on top of standard token pricing, which makes this look less like a pure model application programming interface and more like rented operating infrastructure for agents. (the-decoder.com) There is a catch for big companies with strict cloud rules. The service currently runs only on Anthropic’s own infrastructure, and The Decoder says the launch announcement did not mention support through Amazon Bedrock or Google Vertex Artificial Intelligence. (the-decoder.com) So this launch is not just another model feature. It is Anthropic moving one layer down the stack, from selling intelligence on demand to selling the managed environment where that intelligence actually works all day. (claude.com) (anthropic.com)
