Anthropic opens managed agents beta

Anthropic just made a bigger bet on agents than “here’s a model, go build something.” It opened Claude Managed Agents in public beta on April 8, and the point is simple: developers define the job, the tools, and the guardrails, and Anthropic runs the agent stack for them. That matters because the hard part of production agents usually is not the prompt. It is the messy infrastructure around the prompt — state, permissions, retries, tracing, and secure execution. (claude.com) ### What is Anthropic actually launching? Managed Agents is a hosted service on the Claude Platform for building and deploying cloud-hosted agents at scale. Anthropic is packaging the runtime pieces that teams usually have to assemble themselves — sandboxed code execution, checkpointing, credential handling, scoped permissions, and tracing — into one managed layer. Instead of self-hosting the loop (claude.com)t off to Anthropic. (claude.com) ### Why is that a real product shift? Until now, Anthropic mostly sold model access plus some developer tooling. Managed Agents moves it up the stack. Anthropic is not just saying Claude can power your agent. Anthropic is saying it can run your agent in production. That is a different business. It puts the company closer to the workflow and governance layer that enterprise buyers actually worry about. (claude.com) ### What problem is this trying to fix? Prototype agents are easy to demo. Production agents are annoying. They need to survive disconnects, keep state over long tasks, recover from tool failures, and avoid overbroad access to internal systems. Anthropic’s pitch is that teams burn months building that plumbing before users see anything useful. Managed Agents is supposed to collapse that work into a service so teams can focus on the task design instead. (claude.com) ### What do developers actually control? The key interface is pretty opinionated. Developers define tasks, tools, and guardrails. Anthropic’s orchestration harness decides when to call tools, how to manage context, and how to recover from errors. That means developers get less low-level control than in a fully self-hosted setup, but they also inherit a lot less infrastructure pain. Basically, Anthr(claude.com).” (claude.com) ### What are the standout features? The most important ones are long-running sessions, built-in tracing, secure sandboxing, and governance controls. Sessions can keep running for hours and persist through disconnections. The console shows tool calls, decisions, and failure modes. Anthropic is also previewing multi-agent coordination, where one agent can spin up other agents for parallel work, plus (claude.com)riteria. Some of those more advanced pieces are still in research preview rather than broad beta. (claude.com) ### Why does Anthropic keep talking about “harnesses”? Because the harness is the hidden product. It is the loop around the model — when to call Claude again, when to reset context, when to use tools, when to stop. Anthropic’s engineering write-up says those assumptions go stale as models improve, so it built Managed Agents around stable interfaces like sessions, harnesses, and sandboxes that can c(claude.com)hing. Think operating system abstractions, but for agents. (anthropic.com) ### What is the catch? Convenience trades off against control and lock-in. If Anthropic owns the runtime, Anthropic also defines the abstractions, the debugging surface, and part of the operational model. That will appeal to product teams trying to ship fast. But infra teams may still want self-hosted systems where they can tune every step, swap models freely, or keep more execution insid(anthropic.com)plit. (claude.com) ### Bottom line This launch matters because it turns Anthropic from a model vendor into more of an agent platform company. The headline is not just “Claude can do agentic work.” The headline is “Anthropic wants to run the boring, brittle, expensive parts for you” — and that is exactly where enterprise AI budgets are starting to move. (claude.com)