Agents go mainstream

For the last two years, building an artificial intelligence agent has looked easy in demos and miserable in production. The model could write code or search the web, but the team still had to build the cage, the memory, the logging, and the recovery system around it. (anthropic.com) Anthropic’s bet is that developers no longer want just the brain. On April 8, 2026, it put Claude Managed Agents into public beta as a hosted service that runs long-horizon agents on Anthropic’s own infrastructure. (anthropic.com) A managed agent is basically rented factory machinery instead of a box of parts. You define the agent, the tools, and the guardrails, and Anthropic handles the container, the execution loop, and the live event stream. (platform.claude.com) Anthropic split the system into three pieces so one crash does not kill the whole job. It describes those pieces as the session, which is the running log, the harness, which is the loop that calls Claude, and the sandbox, which is the isolated workspace where code runs and files change. (anthropic.com) That design came from a very old infrastructure lesson. Anthropic says its first version put everything in one container, which turned each agent into a fragile “pet,” so if the container failed, the session was lost with it. (anthropic.com) The new service is also priced to make experimentation feel cheap. Multiple reports published on April 9 and April 10 say Claude Managed Agents costs $0.08 per active session-hour on top of normal token charges, with idle time not billed. (the-decoder.com) (thesys.dev)) (lowcode.agency) That number changes the math for a startup in a hurry. A team that used to spend weeks wiring sandboxing, state management, and retries can now pay cents per hour for the scaffolding and spend its time on the workflow itself. (siliconangle.com) (the-decoder.com) Anthropic is pushing this as more than a toy for hackers. Its documentation says developers can configure packages and network access in the environment, start sessions through the application programming interface, and stream status updates until the agent goes idle. (platform.claude.com) The company also has reference customers ready for the pitch. Anthropic’s customer material says Rakuten is using Claude Managed Agents to ship major releases every two weeks instead of once a quarter, while cutting initial critical errors by 97 percent. (claude.com) This is the bigger shift underneath the product launch: the market is moving from selling intelligence by the token to selling finished labor with infrastructure attached. Anthropic already had strong models for long-running work, including Claude Opus 4.6, and Managed Agents turns that model strength into a hosted operations product. (anthropic.com 1) (anthropic.com 2) If this pricing holds after beta, the first thing that gets cheaper is not inference. It is the cost of trying weird ideas, because a developer can now spin up an agent session like cloud compute instead of building an entire miniature platform before the first test. (platform.claude.com) (the-decoder.com)