Claude Managed Agents Public Beta

Anthropic is trying to turn “AI agents” from a demo that can book a dinner into a service that can keep working after you close your laptop. On April 9, 2026, the company published an engineering post saying Claude Managed Agents is a hosted service for “long-horizon agent work,” and Claude’s official account said the product is now in public beta. (anthropic.com) An agent is just a language model wrapped in a work loop. The model reads a task, calls tools, writes files, checks results, and repeats that cycle until it finishes or fails, which is why Anthropic keeps using the word “harness” for the software around the model. (anthropic.com) That harness is usually the fragile part. Anthropic says developers hard-code assumptions about what the model cannot do, and those assumptions go stale when a newer model suddenly handles the same job without the extra scaffolding. (anthropic.com) Anthropic gives one concrete example from its own testing. Claude Sonnet 4.5 would sometimes stop early as it approached its context limit, so Anthropic added “context resets,” but the same workaround became unnecessary when the company switched that harness to Claude Opus 4.5. (anthropic.com) Managed Agents is Anthropic’s answer to that moving-target problem. Instead of asking every customer to build and babysit their own agent runtime, Anthropic says it will run the agent on Claude Platform through a small set of interfaces that are supposed to survive even when the machinery underneath changes. (anthropic.com) The company breaks that machinery into three parts. A session is the append-only log of everything that happened, a harness is the loop that calls Claude and routes tool calls, and a sandbox is the execution environment where Claude can run code and edit files. (anthropic.com) Anthropic compares that design to an operating system. A program can still use the same file command whether the storage is a disk from the 1970s or a modern solid-state drive, and Anthropic says Managed Agents is trying to give developers that same kind of stable layer for agent workflows. (anthropic.com) That sounds abstract until you hit production failures. Anthropic says its first version put the session, harness, and sandbox in one container, which made the server a “pet” instead of “cattle,” and a single container failure could wipe out the whole session. (anthropic.com) The timing also fits Anthropic’s product push over the last year. The company introduced Claude 4 in May 2025 as a model family built for sustained agent workflows, added advanced tool use in November 2025 so Claude could discover and execute tools dynamically, and upgraded Claude Opus 4.6 in February 2026 with stronger performance on long-running agentic tasks. (anthropic.com 1) (anthropic.com 2) (anthropic.com 3) What changed this week is the packaging. Anthropic is no longer just saying “our models are good at agents”; it is selling the surrounding runtime as a managed product, which puts it closer to cloud software you deploy with controls than to a chatbot feature you experiment with on weekends. (anthropic.com)