Claude Managed Agents add sandboxing, $0.08/hr

Anthropic has moved further from “here’s a model API” and closer to “here’s the whole agent runtime.” Claude Managed Agents entered public beta on April 8, 2026, and the pitch is simple: Anthropic hosts the loop, the tools, the session state, and the sandbox so developers do not have to wire that stack together themselves. Then, on April 23, Anthropic added memory in public beta, which makes the product feel more like a real long-running worker than a stateless chatbot. (platform.claude.com) ### What is Anthropic actually selling? Basically, Managed Agents is a hosted harness for running Claude as an autonomous agent. Anthropic’s docs frame it as the alternative to the lower-level Messages API: instead of building your own agent loop, tool execution layer, and runtime, you create an agent, create an environment, start a session, and stream events back(platform.claude.com)-shot prompting. (platform.claude.com) ### Why does the sandbox matter? Because “agent” demos are easy, but safe execution is the annoying part. Anthropic says the managed setup gives Claude secure access to built-in tools like bash, file operations, web search and fetch, plus code execution inside cloud containers you configure. The environment can include preinstalled packages and network rules, whi(platform.claude.com)own. (platform.claude.com) ### What changed in April? Two things. First, Anthropic launched Claude Managed Agents in public beta on April 8 under a dedicated beta header. Second, on April 23, it added memory for Managed Agents in public beta. That second step matters because it lets an agent carry information across sessions instead of starting blank every time. Anthropic’s examples are user preferences, project conventions, prior mistakes, and domain context. (platform.claude.com) ### How does pricing work? There are two meters. You still pay for model tokens, and Managed Agents adds session runtime pricing on top. Anthropic’s pricing docs say Managed Agents are billed on tokens plus session runtime, and the listed runtime price is $0.08 per session-hour. Anthropic also notes that this replaces the older container-hour billing model for cod(platform.claude.com)top of session runtime. (platform.claude.com) ### Why not just use the Messages API? You still would if you want tight control. Anthropic’s own docs split the products pretty cleanly: Messages API is for direct prompting, custom loops, and fine-grained orchestration; Managed Agents is for stateful sessions and hosted infrastructure. So this is less a replacement than a vertical move up the stack. Anthropic wants to sell not just intelligence, but the scaffolding around intelligence. (platform.claude.com) ### Who is this really for? Teams building internal automation, coding agents, research workflows, and other jobs that can run for minutes or hours. The docs keep emphasizing persistent event history, server-side state, secure containers, and minimal infrastructure. That is enterprise language for “you can skip building the backend plumbing.” If your company has (platform.claude.com)e managed version of that pile. (platform.claude.com) ### What is the catch? The catch is that this is still beta, and it lives behind a specific beta header. Also, hosted agent runtimes can simplify architecture while making you more dependent on one vendor’s abstractions, pricing model, and tool surface. That tradeoff is fine for a lot of companies — but it is still a tradeoff. (platform.claude.com)ng to make agent deployment feel boring in the best way. Not magical — just managed. The notable part is not only sandboxing or memory or the $0.08 session-hour price by itself. It is the bundle. Anthropic is packaging model access, runtime, tools, and state into one product, which pushes the company deeper into platform territory. (platform.claude.com)