Claude launches Managed Agents
Claude announced Managed Agents in public beta, moving beyond short chat replies toward agents that run for hours and manage longer automation workflows. (x.com) Media coverage highlights these agents as long‑running workers for engineering tasks—planner/executor flows, checkpointing, tool calls and human approval gates—rather than simple copilots. (youtube.com)
Most people have used artificial intelligence as a fast typist: ask a question, get a paragraph, move on. Anthropic’s new Managed Agents product is built for a different job entirely: software workers that stay on task for hours inside Anthropic’s own cloud instead of replying once and disappearing. (claude.com) That shift solves a specific engineering problem. A one-shot chatbot only needs a prompt and a reply, but a long job like “debug this service, run tests, edit files, and come back for approval” also needs memory, permissions, logging, and a safe place to run code. (claude.com) Anthropic says Managed Agents is a public beta “suite of composable application programming interfaces,” which is developer language for building blocks you can snap together. The company says developers set the task, tools, and guardrails, while Anthropic handles sandboxed code execution, checkpointing, credential management, scoped permissions, and tracing. (claude.com) A sandbox is just a sealed workroom. It lets the model run code and edit files without being dropped loose onto a production machine the way a junior engineer would never be handed the keys to every server on day one. (anthropic.com) A checkpoint is a save point. Anthropic already added checkpoints to Claude Code in September 2025 so users could rewind code and conversation to an earlier state after a risky change, and Managed Agents carries that same “try something, then roll back” idea into longer workflows. (anthropic.com) The hard part in these systems is not only intelligence but continuity. Anthropic wrote in November 2025 that long-running agents break work into separate sessions, and each new session can start “with no memory of what came before” unless the system leaves behind clear artifacts for the next pass. (anthropic.com) That is why Anthropic keeps talking about a harness. A harness is the control loop around the model that decides when to call tools, how to manage context, and how to recover after an error, which makes it closer to an operating system for an agent than a simple chat box. (claude.com) Anthropic’s engineering team says older harnesses often baked in assumptions that went stale as models improved. In one example, Claude Sonnet 4.5 needed “context resets” because it would wrap up tasks too early near its context limit, while Claude Opus 4.5 no longer showed that behavior, so the old fix became extra baggage. (anthropic.com) Managed Agents is Anthropic’s answer to that moving target. The company says it split the system into stable parts — a session, which is the running log; a harness, which is the control loop; and a sandbox, which is the execution room — so Anthropic can swap the machinery underneath without forcing customers to rebuild their apps every time the models change. (anthropic.com) The timing also lines up with how people are already using these tools. Anthropic said in February 2026 that among the longest Claude Code sessions, autonomous work time had nearly doubled in three months from under 25 minutes to over 45 minutes, and software engineering made up nearly 50% of agent activity on its public application programming interface. (anthropic.com) So this launch is less about a smarter chatbot than about Anthropic trying to own the plumbing for autonomous work. If developers accept Anthropic’s hosted sessions, guardrails, and recovery systems, the company moves one layer up the stack from selling model access to running the whole job. (claude.com)
