Cursor Launches Long-Running Autonomous Agents
AI-native code editor Cursor has launched "long-running agents" for its Ultra+ users, enabling autonomous execution of tasks that can last for hours or days. These agents can handle large-scale code refactoring and complex DevOps workflows without continuous human supervision. The development is seen as a move toward "action engines," where AI transitions from suggesting to executing end-to-end tasks with persistent memory.
- The "long-running" capability is designed for tasks lasting between 25 and 52 hours without requiring human supervision, a significant leap from previous agents that operated in shorter prompt-response loops. This feature is available exclusively for users on Cursor's Ultra, Teams, and Enterprise plans. - Cursor's parent company, Anysphere, was founded in 2022 by four MIT graduates: Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger. By November 2025, a $2.3 billion funding round valued the company at $29.3 billion, making the co-founders billionaires. - The development of these agents involved creating a custom "harness" to overcome the limitations of frontier models on long-horizon tasks, such as losing track of the big picture or stopping at partial completion. This system separates tasks between "Planner" agents that create tasks and "Worker" agents that focus solely on execution. - Early applications of the long-running agents during internal testing included refactoring an authentication system (25-hour runtime), building a new mobile app based on an existing web app (30-hour runtime), and migrating a video renderer to Rust for a 25x performance improvement. - The company's valuation has grown rapidly, reaching a reported $9.9 billion after a $900 million funding round in June 2025 and previously hitting $2.5 billion in January 2025. Key investors include Andreessen Horowitz, Thrive Capital, Accel, and the OpenAI Startup Fund. - The competitive landscape for autonomous coding agents includes tools like Windsurf (formerly Codeium), GitHub Copilot, and the terminal-based Claude Code. Alternatives are often distinguished by price, workflow (interactive vs. fully autonomous), and whether they are open-source. - This technology is part of a broader shift from "AI copilots," which assist humans, to "AI agents," which can act autonomously. Agentic AI is characterized by its ability to plan, use tools, and maintain memory to achieve goals with minimal human intervention. - To test the scaling limits of their agent architecture, Cursor ran an experiment to build a functional web browser from scratch, which involved hundreds of concurrent agents writing over 1 million lines of code in about a week.
