Anthropic ships managed agents

Anthropic just started selling the part of artificial intelligence that most companies never show in demos: the scaffolding behind the bot. On April 8, 2026, it opened Claude Managed Agents in public beta, promising cloud-hosted agents that companies can deploy without wiring up the whole runtime themselves. (claude.com) An artificial intelligence agent is a model that does more than answer one prompt. It keeps working across many steps, uses tools, checks results, and hands back a finished job more like a contractor than a chatbot. (platform.claude.com) That sounds simple until you try to run one in production. A company has to keep the agent’s memory straight, decide which tools it can touch, give it a safe computer to work inside, and make sure it does not wander into the wrong system. (anthropic.com) Anthropic’s pitch is that those moving parts should be rented, not rebuilt. Its new service bundles the agent, the environment it runs in, the session that tracks a job, and the event stream that reports what happened while the work is in progress. (platform.claude.com) The “environment” is basically a preconfigured container, which is a sealed-off workspace with approved packages and network rules. The “session” is one live run of the agent, so a company can start a task, watch it work, and collect outputs without building its own control room first. (platform.claude.com) Anthropic says this can cut the trip from prototype to production by a factor of 10. In its launch post, the company framed Managed Agents as “composable application programming interfaces,” meaning developers can snap pieces together instead of writing one giant custom system. (claude.com) The company’s engineering team says the old way breaks whenever models improve. Their argument is that hard-coded “harnesses,” which are the brittle scripts wrapped around a model, age fast, so Anthropic wants stable interfaces even as the model underneath gets smarter. (anthropic.com) This launch also fits the way Anthropic has been positioning Claude through 2026. Its Claude Opus 4.6 model page says the model is built for “complex multi-agent work,” and its broader developer docs already offer an Agent Software Development Kit for teams that still want to run more of the loop themselves. (anthropic.com, platform.claude.com) So Anthropic is now selling both layers at once. One layer is the brain, meaning the model; the other layer is the managed workplace, meaning the hosted system that lets the model use tools and finish long jobs inside guardrails. (anthropic.com, claude.com) The timing is not subtle. Wired reported that Anthropic has been touting rapid revenue growth alongside the launch, and Bloomberg reported on April 8 that the company said it had passed $30 billion in annualized run-rate revenue in April after topping $19 billion in March. (wired.com, bloomberg.com) That makes Managed Agents less like a side feature and more like a bet on where the money is going. If companies stop paying only for tokens and start paying for the full operating layer around autonomous work, Anthropic gets closer to being the landlord as well as the model maker. (claude.com, wired.com)