Anthropic splits Claude into three

Anthropic is now marketing Claude as three different products: Claude for chat and research, Claude Code for software engineering, and Claude Cowork for desktop knowledge work. (anthropic.com 1) (anthropic.com 2) The split is visible across Anthropic’s own product pages and research posts, which describe Claude Code as an “agentic coding system” and Cowork as software that executes multi-step office tasks on a user’s computer. (anthropic.com 1) (anthropic.com 2) Claude Code is built to read a codebase, edit files, run tests, use command-line tools, and deliver committed code. Anthropic says Stripe rolled it out to 1,370 engineers, and one team used it for a 10,000-line Scala-to-Java migration in four days. (anthropic.com) Claude Cowork is the non-engineering counterpart. Anthropic says it can work across local files, folders, and desktop apps to organize documents, extract data from dense files, and prepare reports from source material. (anthropic.com) Anthropic’s April 9 research note frames both products as part of the same shift from chatbot use to “agents” that plan, act, observe results, and loop until a task is done or needs human input. (anthropic.com) That matters because Anthropic is also building the infrastructure behind longer-running agents. In an engineering post published in April, the company said its Managed Agents service separates an agent into three stable abstractions: session, harness, and sandbox. (anthropic.com) A session is the running log of what happened, a harness is the loop that calls Claude and routes tool use, and a sandbox is the execution environment where Claude can run code and edit files. Anthropic says those interfaces are meant to stay stable even as the underlying implementation changes. (anthropic.com) The company’s own explanation for that design is that agent “harnesses” age quickly as models improve. Anthropic says fixes it needed for Claude Sonnet 4.5, including context resets to stop premature wrap-ups near token limits, became unnecessary with Claude Opus 4.5. (anthropic.com) Anthropic has been arguing for months that building useful agents is less about a single prompt than about managing the full working context over many turns. Its context-engineering post calls that process the curation of instructions, tools, external data, and message history inside a limited context window. (anthropic.com) The result is a cleaner product map: one Claude for asking questions, one for building software, and one for handing off repetitive desktop work. Anthropic is selling not just a model, but different operating surfaces for different jobs. (anthropic.com)