Anthropic packages enterprise control

Anthropic just moved one layer up the artificial intelligence stack: instead of only selling access to Claude models, it is now selling the operating system around them, with Claude Managed Agents entering public beta on April 8, 2026. The service runs agents for customers inside Anthropic’s cloud and adds secure sandboxing, built-in tools, and live event streaming over the application programming interface. (anthropic.com) That solves a boring problem that has slowed down most “agent” projects inside companies. A model can write a plan in seconds, but turning that plan into software that opens files, calls tools, stores state, and survives errors usually takes weeks or months of engineering work. (wired.com) Anthropic’s own engineering team describes this missing layer as the “harness.” A harness is the wrapper that gives a model hands and guardrails at the same time, and Anthropic says those wrappers keep breaking as models improve because old assumptions go stale. (anthropic.com) Claude Managed Agents is Anthropic’s answer to that harness problem. The company says customers can create agents, configure containers, and run sessions through the application programming interface, while Anthropic handles the runtime pieces like sandboxing and session management. (anthropic.com) The security pitch is simple: keep the agent in a box. VentureBeat reported this week that enterprise security teams are fixated on where credentials live and how far an agent can reach after a mistake, and Anthropic’s architecture is designed to limit that blast radius with stronger isolation. (venturebeat.com) Anthropic paired that developer product with a second launch aimed at office workers instead of engineers. On April 9, 2026, it took Claude Cowork out of preview and made it generally available on paid Claude plans including Pro, Team, and Enterprise. (thenewstack.io) Claude Cowork is not a chat window that waits for one prompt at a time. Anthropic describes it as a desktop agent for multi-step knowledge work like research synthesis, document preparation, file management, and work across local folders and everyday applications. (anthropic.com) The enterprise upgrade is mostly about who gets to touch what. Anthropic added role-based access controls, group spend limits, usage analytics, OpenTelemetry monitoring, and per-tool connector controls so an information technology team can decide which department can use Claude Cowork and which tools each group can reach. (support.claude.com) Those per-tool controls sit on top of the Model Context Protocol, which is Anthropic’s open standard for connecting models to outside tools and data sources. Anthropic donated that protocol to the Linux Foundation-backed Agentic AI Foundation in December 2025 with support from companies including OpenAI, Google, Microsoft, Amazon Web Services, Cloudflare, and Bloomberg. (anthropic.com) So the company is now selling both halves of the same promise. Managed Agents gives developers a hosted control plane for production agents, while Claude Cowork gives business teams a governed desktop worker, and both products are wrapped in the same language of permissions, observability, and policy. (siliconangle.com (anthropic.com)) That is a shift in what the competition is about. The fight is no longer just whose model writes the best paragraph or code snippet; it is who can make autonomous software acceptable to a security team, an auditor, and a procurement department. (siliconangle.com)