Anthropic launches Managed Agents
Anthropic has rolled out Claude Managed Agents — a hosted, production-ready way to build and run multi-agent workflows so teams can move from prototypes into production faster. The launch positions orchestration, runtime and hosting as a product rather than just an API, and was discussed publicly alongside social posts highlighting the move into packaged agent deployments. (winbuzzer.com / x.com)
Most companies can get a chatbot demo working in a weekend. The part that eats the next three months is everything around it: secure code execution, saved state, permissions, tracing, and recovery when a tool call fails. Anthropic’s new product is aimed at that second part, not the demo. (claude.com) Anthropic launched Claude Managed Agents in public beta on April 8, 2026, on the Claude Platform. Instead of giving developers just a model endpoint, Anthropic is now offering a hosted runtime where Claude can keep working on a task for minutes or hours. (claude.com, platform.claude.com) An agent is just a language model with tools and a to-do list. The hard part is the loop that decides when to search the web, when to run code, when to read a file, and how to pick back up after something breaks. (platform.claude.com, anthropic.com) Anthropic says its managed setup gives Claude built-in access to shell commands, file operations, web search, web fetch, and Model Context Protocol servers, which are connectors to outside tools. A developer creates the agent once, creates a cloud environment once, and then starts sessions inside that setup. (platform.claude.com) Those sessions are the key unit here. Anthropic describes a session as a running agent instance with server-side event history, so a team can interrupt it, steer it, and resume work without rebuilding the whole run from scratch. (platform.claude.com) Under the hood, Anthropic split the system into three pieces: the session, the harness, and the sandbox. The session is the durable log, the harness is the traffic cop that routes model and tool calls, and the sandbox is the locked room where code runs and files get edited. (anthropic.com) That split fixes a boring but expensive problem. Anthropic says its earlier design put everything in one container, and when that container failed, the session was lost, which turned the server into a fragile “pet” instead of a replaceable machine. (anthropic.com) Anthropic is also selling this as a speed play. Its launch post says teams can get to production “10x faster” and move from prototype to launch in days rather than months because Anthropic handles sandboxing, checkpointing, credential management, scoped permissions, and tracing. (claude.com) The product is not just for one-shot prompts. Anthropic’s docs say Managed Agents is meant for long-running and asynchronous work, with persistent file systems, conversation history across interactions, and cloud containers that can be preloaded with languages like Python, Node.js, and Go. (platform.claude.com) Anthropic is already naming customers to show this is past the lab stage. Its launch materials and follow-on coverage say Notion, Asana, and Sentry have used Managed Agents, while 9to5Mac reported the service is now available on the Claude Platform for all builders. (claude.com, 9to5mac.com) The bigger shift is what Anthropic is selling. For the last two years, most artificial intelligence companies sold brains through application programming interfaces; Anthropic is now packaging the brain, the hands, and the workplace into one hosted product. (anthropic.com, claude.com)
