Anthropic sells hosted agent runtime

Anthropic just moved one layer down the stack: instead of only selling the brain, it is now selling the place where the brain lives and works. On April 8, 2026, it launched Claude Managed Agents in public beta, a hosted service for running autonomous agents on Anthropic’s own infrastructure. (the-decoder.com) That sounds abstract, but the problem is concrete. A company can get a demo agent working in a day, then spend months building the boring parts around it: sandboxes, session memory, credentials, recovery after crashes, and the loop that keeps a task going for hours instead of one reply. (wired.com) Anthropic’s pitch is that developers should define the job, the tools, and the guardrails, and Anthropic should run the machinery underneath. The service is priced at normal Claude token rates plus $0.08 per session hour, which turns agent runtime into a metered product instead of a pile of custom code. (the-decoder.com) The early users show what Anthropic wants this to become. The Decoder reports that Notion is using it for workspace delegation, Rakuten built internal agents for sales, marketing, and finance inside Slack and Microsoft Teams, and Sentry paired it with a debugging flow that can write patches and open pull requests. (the-decoder.com) There is a catch in the launch: for now, these agents run only on Anthropic’s infrastructure. The announcement did not say the service would also be available through Amazon Bedrock or Google Cloud Vertex AI, which matters for big companies that spread workloads across multiple clouds. (the-decoder.com) A day later, on April 9, 2026, Anthropic pushed Claude Cowork out of research preview and into general availability on paid Claude plans. Claude Cowork is the desktop product for non-technical workers who want an agent to handle multi-step jobs with files, folders, and everyday apps instead of chatting turn by turn. (thenewstack.io) (anthropic.com) The enterprise features tell you who this launch is for. Anthropic added role-based access controls, group-level spend caps, usage analytics, OpenTelemetry support for monitoring, and admin controls over connectors and tools, which is the checklist information technology departments ask for before they let software touch company systems. (thenewstack.io) (anthropic.com) Anthropic is stitching those two products together from opposite ends. Managed Agents is for developers building agent systems through an application programming interface, while Claude Cowork is for employees delegating work through a desktop app, and both depend on Anthropic owning the runtime, the permissions, and the audit trail. (the-decoder.com) (anthropic.com) That is a different business from selling model access alone. A model answers a prompt; a runtime keeps an agent alive, stores its state, decides what tools it can touch, and records what happened when it acted inside a company. (wired.com) (the-decoder.com) Anthropic has been hinting at this direction for months in its engineering work on long-running agents and harness design, and now it has productized that work for customers. If this sticks, the competition shifts from “whose model is smartest” to “whose managed environment is trusted enough to let an agent actually do the job.” (anthropic.com) (thenewstack.io)