Claude Managed Agents go broadly available with $0.08 per session-hour pricing and secure sandboxing

AI agents are getting less like chat windows and more like remote workers. That sounds like marketing fluff, but this release is really about infrastructure — the boring, expensive part that usu(platform.claude.com)ged Agents are now available as a hosted service for long-running work, with secure execution containers, built-in tools, and sessions that persis(wowhow.cloud)matter: you still pay normal Claude token rates, and then add $0.08 per session-hour for the runtime itself. (platform.claude.com)ually selling? Managed Agents is basically Anthropic’s answer to the “I don’t want to build my own agent runtime” problem. Instead of wiring together your own loop, sandbox, tool runner, and state management, you define an agent, define an environment, start a session, and let Claude operate inside that managed setup. Anthropic handles the harness and infrastructure around the model. (platform.claude.com) ### Why is that different from just calling a model API? A plain model API gives you responses. This gives you an agent that can keep working for minutes or hours, call tool(platform.claude.com)k while the session state stays persisted server-side. That matters because most production agent failures are not “the model answered badly.” They’re “the process crashed,” “the tool chain broke,” or “we lost the task state halfway through.” (platform.claude.com) ### What does the secure sandbox part mean? Claude runs actions inside managed cloud (platform.claude.com)ngineering writeup describes this as separating the “brain” from the “hands” — the model reasons, while the execution happens in a sandbox that can be swapped or recovered without taking the whole session down. In plain English, the agent can do real work without being fused to one fragile machine. (platform.claude.com) ### Where does memory fit in? By default, each session still starts fresh. But Anthropic has added memory stores in (platform.claude.com)nts mounted into the session like files, so the agent can read and write them using normal file tools. Every change creates an immutable version, which gives teams an audit trail and point-in-time recovery. That’s a big shift from “the agent remembers stuff somehow” to “the memory is inspectable, editable, exportable files.” (platform.claude.com) ### Why does the $0.08 number matter so much? Because it tur(platform.claude.com)ated agent costs as mostly token costs plus whatever hidden cloud mess they built themselves. Managed Agents makes the tradeoff explicit: pay Anthropic for the runtime layer and skip a chunk of custom engineering. Secondary guides citing Anthropic’s pricing page peg that runtime fee at $0.08 per session-hour on top of token charges. (wowhow.cloud) ### Is this just a developer convenience feature? Not really. Anthropic has been making a (platform.claude.com)ey move from toy tasks into coding, finance, healthcare, and security-adjacent work. Its recent research on trustworthy agents and agent autonomy is basically the backdrop here — more autonomous systems need better oversight, clearer logs, and safer execution boundaries. Managed Agents looks like the product version of that thesis. (anthropic.com) ### So what changed this week? The important ch(wowhow.cloud) that architecture into a product developers can use now, and then layered persistent memory onto it in beta. That moves Claude further from “smart assistant” and closer to “managed worker process with a model attached.” (platform.claude.com) ### Bottom line The real news is not that Claude got smarter. It’s that Anthropic is trying to own the runtime around smart models — the sandbox, the session, the logs, and now the memory. If agents become durable software workers, that layer could end up being the business. (platform.claude.com)