‘Conway’ leak hints at persistent Claude

A leak on X describes an internal Anthropic project called “Conway” as an always-on version of Claude that can keep running, use tools, and automate browser work after the first prompt is over. The posts are not confirmed by Anthropic, but the details line up with capabilities the company has already been shipping in public over the past 18 months. (x.com) (anthropic.com) Most people still use chatbots like a calculator with amnesia: you ask a question, get an answer, and start over. A persistent agent works more like an employee with a desk, a to-do list, and permission to open software, revisit tasks, and keep going while you are away. (anthropic.com) That difference sounds small until you look at how work actually happens on a computer. Real tasks usually spill across tabs, files, forms, calendars, dashboards, and login screens, which means the useful system is not the one that writes the best paragraph, but the one that can stay oriented for dozens or hundreds of steps. (anthropic.com 1) (anthropic.com 2) Anthropic has been moving toward that model in public for a while. In October 2024, it introduced “computer use” for Claude 3.5 Sonnet, letting developers direct Claude to look at a screen, move a cursor, click buttons, and type into a virtual keyboard. (anthropic.com 1) (anthropic.com 2) That first release mattered because it pushed Claude beyond text generation and into software control. Anthropic said companies including Asana, Canva, Cognition, DoorDash, Replit, and The Browser Company were already testing workflows that required dozens or even hundreds of actions. (anthropic.com) By late 2025, Anthropic was talking much more openly about “long-running agents.” In engineering posts, the company described agent harnesses that let Claude plan, use tools, compact context, recover from errors, and continue useful work across multiple context windows instead of stopping when one chat gets too long. (anthropic.com) Anthropic also kept adding the plumbing that persistent agents need. In November 2025, it announced advanced tool use features that let Claude discover, learn, and execute tools dynamically, which is a step away from hand-wired single-purpose chatbots and toward systems that can improvise across large toolsets. (anthropic.com) The company’s own product lineup now reflects that shift. Anthropic’s site in April 2026 highlights “AI agents,” “Claude Code,” “Claude Cowork,” “Claude for Chrome,” and enterprise workflows, while current model pages for Claude Sonnet 4.6 and Claude Opus 4.6 explicitly pitch “long-running agents,” “complex agentic workflows,” and professional tasks that extend beyond one session. (anthropic.com 1) (anthropic.com 2) (anthropic.com 3) If the Conway leak is real, the new piece is not tool use by itself. The new piece is persistence: a Claude variant that stays available, remembers enough state to keep acting, and treats the browser and other software as places to work rather than websites to summarize. (x.com) (anthropic.com) That would change how people build workflows around Claude. Instead of writing giant prompts that try to predict every step in advance, developers could assign goals, connect calendars or customer systems or internal dashboards, and let the model return only when a job is finished or blocked. (anthropic.com) (anthropic.com) It would also raise the same safety problem Anthropic has been warning about in its own research. A model that reads the open web and clicks through software can be tricked by prompt injection, where hidden instructions inside pages or documents try to hijack the agent’s behavior; Anthropic said in November 2025 that no browser agent is immune to that risk. (anthropic.com) So the leak does not read like a sudden left turn. It reads like a private name for a direction Anthropic has already been signaling in public: Claude moving from a chat window you visit to a system that can keep working in the background, across tools, over time. (x.com) (anthropic.com) (anthropic.com) Anthropic has not publicly confirmed a product called Conway as of April 8, 2026, so the safest reading is that the leak hints at an experiment, not a launch. But even as an experiment, it fits the company’s recent releases closely enough to show where the next fight in artificial intelligence products is headed: not better answers in one turn, but software that can keep acting after you close the tab. (x.com) (anthropic.com)