Anthropic ships Managed Agents

Anthropic is trying to sell companies the part of artificial intelligence that usually breaks after the demo: the plumbing that keeps an agent running when it has to use tools, remember state, and recover from errors for hours instead of minutes. On April 8, 2026, it launched Claude Managed Agents in public beta as a hosted service on the Claude platform. (claude.com) An agent is just a language model with hands: it can call tools, write files, run code, and take multiple steps toward a goal instead of answering once and stopping. Anthropic says the hard part is not the model itself but the production work around sandboxed code execution, checkpointing, credential management, scoped permissions, and tracing. (claude.com) Managed Agents is Anthropic’s answer to that problem. The company says customers define the task, tools, and guardrails, and Anthropic runs the agent on its own infrastructure instead of making each customer build the control system from scratch. (claude.com) Anthropic’s engineers describe that control system as a harness, which is the loop that decides when the model should think again, call a tool, or recover after something fails. In its engineering write-up, Anthropic says old harnesses can become dead weight when newer models stop needing the same workarounds, so it wants stable interfaces above changing internals. (anthropic.com) The company says it broke the service into three pieces: a session, which is the running log of what happened; a harness, which routes model and tool calls; and a sandbox, which is the isolated workspace where code runs and files get edited. Anthropic compares that design to an operating system keeping the same file and process commands even as the hardware underneath changes. (anthropic.com) That pitch is aimed at businesses that like agent demos but do not want to babysit fragile containers and custom orchestration code. Anthropic says Managed Agents can get teams to production “10x faster,” while outside coverage says the product is meant to cut development timelines from months to weeks or even days. (claude.com) (siliconangle.com) (testingcatalog.com) Anthropic paired that launch with a second enterprise push around Claude Cowork, its macOS app for working alongside Claude on desktop tasks. On April 9, 2026, 9to5Mac reported that Claude Cowork was dropping its “research preview” label and adding enterprise capabilities at the same time Anthropic rolled out Managed Agents. (9to5mac.com) At the same moment Anthropic is asking companies to trust it with more automation, it is also drawing a hard line around a separate model called Mythos. TechCrunch reported on April 7, 2026 that Anthropic released Mythos only to a small set of partners in Project Glasswing and said the preview would not be generally available. (techcrunch.com) Anthropic told TechCrunch that Mythos had identified thousands of zero-day vulnerabilities over the previous few weeks, many of them critical and many buried in software that was one to two decades old. The partner list included Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks, which shows Anthropic is treating Mythos more like a restricted security instrument than a normal product launch. (techcrunch.com) So the company shipped two messages in one week. Managed Agents says Anthropic wants Claude to move from clever assistant to managed business system, while the Mythos restrictions say Anthropic still believes some model capabilities should stay behind locked doors even when they are commercially valuable. (claude.com) (techcrunch.com)