Anthropic publishes managed‑agent docs
Anthropic released formal documentation for Claude Managed Agents, describing a configurable agent harness for long‑running, asynchronous tasks that handles retries, state and orchestration so developers don't have to build that plumbing themselves. The docs also map model choices across the Claude family to help balance capability, latency and cost for prototypes. Anthropic has rolled Claude into an enterprise Word integration in beta that can read and edit documents, track comments and assist with contract review — a use case highlighted for legal workflows. (platform.claude.com) (platform.claude.com) (thenextweb.com)
Anthropic has published formal documentation for Claude Managed Agents, a hosted system for running long, multi-step artificial intelligence jobs without building the control layer yourself. (platform.claude.com) The new docs say Managed Agents provides the “harness and infrastructure” for autonomous work, replacing a custom agent loop, tool execution stack, and runtime with Anthropic’s managed service. Anthropic says the agent can read files, run commands, browse the web, and execute code in a secure environment. (platform.claude.com) Anthropic’s quickstart shows the product as a session-based API: developers create an agent and environment, open a session, send user events, and stream updates back while the task runs. The same guide says the service provisions a container, runs the agent loop, executes tool calls, and idles until the next event. (platform.claude.com) An agent is a language model wrapped in software that can take actions over time, not just answer one prompt. Anthropic’s engineering team wrote on April 9 that “harnesses” for these systems often break as models improve, so the company is trying to keep the interface stable while it changes the underlying orchestration. (anthropic.com) That pitch lands at a moment when developers are moving from chatbots to longer-running software workers that need memory, retries, and tool use. Anthropic had already been writing about “effective harnesses for long-running agents” in November 2025 and added more agent-building features to its application programming interface this month, including code execution, file storage, and Model Context Protocol server connections. (anthropic.com) (claude.com) Anthropic also paired the release with model-selection guidance in its Claude documentation. The models page says Claude Sonnet 4.5 is the company’s “smartest model” for complex agents and coding, while Claude Haiku 4.5 is positioned as the faster option for lower-latency work. (platform.claude.com) The company is also pushing Claude deeper into office software. Anthropic’s product pages now list Claude for Word alongside Claude for Excel and Claude for PowerPoint, and a new Skills post says Claude can read and generate Word documents, spreadsheets, presentations, and fillable Portable Document Format files. (anthropic.com) (claude.com) Trade publication Artificial Lawyer reported on April 11 that Anthropic had released Claude for Word in beta, with native tracked changes, comment handling, and legal contract review featured as an example workflow. The report said Anthropic was explicitly targeting lawyers, a customer group that already spends much of its day inside Microsoft Word. (artificiallawyer.com) Taken together, the new agent docs and the Word rollout show Anthropic working both sides of the market at once: developer infrastructure on one side, end-user software on the other. The near-term test is whether companies want Anthropic to run the plumbing for agent systems, or still prefer to assemble that stack themselves. (platform.claude.com) (anthropic.com)
