Anthropic expands both software and hardware horizons

Anthropic is trying to sell companies more than a chatbot now. On April 8, 2026, it put Claude Managed Agents into public beta, a hosted service that runs long, multi-step agents for customers instead of making them wire the whole system together themselves. (anthropic.com) That sounds small until you look at what companies usually have to build first. Anthropic says production agents need things like sessions, tool routing, secure sandboxes, and failure recovery, which is the plumbing that turns a demo into software a business can trust. (anthropic.com) A session is the running memory of a job, like a case file that keeps every step in order. Anthropic says its managed service keeps that log separate from the execution environment so the agent can keep going even if one container dies. (anthropic.com) A harness is the control loop that decides when the model should think, when it should call a tool, and where the result should go next. Anthropic says those harnesses age fast as models improve, so Managed Agents is built around interfaces that stay stable even when the control logic changes underneath. (anthropic.com) This product did not appear out of nowhere. In May 2025, Anthropic added code execution, Model Context Protocol connectors, files, and longer prompt caching to its application programming interface, and in November 2025 it added tool search and programmatic tool calling so agents could work across large tool libraries without stuffing every instruction into memory at once. (anthropic.com, anthropic.com) Those details matter because enterprise agents break in boring ways before they fail in dramatic ways. Anthropic’s November 2025 engineering post said just five connected services could eat roughly 55,000 tokens before a task even starts, which is like filling half a suitcase with instruction manuals before packing any clothes. (anthropic.com) Anthropic is also pushing Claude into big-company workflows where that managed setup is easier to sell. By November 2025, Microsoft had put Claude models into Microsoft Foundry and Microsoft 365 Copilot, and Anthropic had expanded a Salesforce partnership for regulated industries in October 2025. (anthropic.com, anthropic.com) At the same time, Anthropic is looking further down the stack. Reuters reported on April 9, 2026, that the company is exploring designing its own artificial intelligence chips, though the plan is still early and it could decide to keep buying chips instead. (reuters.com via Yahoo Finance) That chip story sits next to a different April 2026 fact: Bloomberg reported on April 6 that Anthropic had topped a $30 billion revenue run rate and expanded a partnership with Broadcom and Google for about 3.5 gigawatts of computing capacity starting in 2027. Anthropic is already securing outside supply at huge scale even while it studies whether owning more of the hardware could lower risk later. (bloomberg.com) So the company is moving in two directions that meet in the middle. It wants to be the service layer that runs agents for businesses today, and it wants the option to control more of the silicon underneath those agents tomorrow. (anthropic.com, reuters.com via Yahoo Finance)