AI-assisted programming is undergoing a fundamental paradigm shift. Recently, Michael Truell, co-founder of Cursor, stated on platform X that the company has officially entered the "third era" of AI programming. The core of this new era is driven by cloud agents capable of independently handling complex tasks and operating autonomously over extended periods. This signifies a fundamental shift in Cursor's positioning, evolving from a "tool for writing code" into a "platform that helps developers build software factories." Data supports this assessment: currently, 35% of internally merged pull requests at Cursor are completed by autonomous agents running in cloud virtual machines. More notably, the company anticipates that the vast majority of software development work will be handled by such agents within a year. This trend is set to reshape the competitive landscape of the AI programming tool sector and profoundly impact the industry's business models.

The evolution speed of AI programming is exceeding all expectations. A key inflection point for the field was reviewed, clearly dividing the development of AI programming tools into three distinct phases. The first phase was the era of Tab autocompletion. The Tab key could not only complete the current line but also intelligently predict and generate subsequent lines of code, even making multi-file modifications, freeing developers from repetitive coding tasks. This phase lasted nearly two years, with the core objective being the automation of low-entropy, repetitive work. The second phase is the era of synchronous agents. The core feature of this era is conversational programming, where developers describe requirements to an agent in natural language, and the agent generates code and responds in real-time, creating a rapid "prompt-feedback-revision" interaction loop. It is predicted that this phase may last less than a year, with the pace of transition far exceeding prior expectations. The third phase is the era of cloud agents. After a developer delegates a task, the agent operates independently within a cloud virtual machine—autonomously handling coding, debugging, testing, and iteration. The developer's role shifts from being a "code writer" to a "director of agents."

User behavior data confirms the intensity of this paradigm shift. In March 2025, the number of Tab users on Cursor was 2.5 times greater than agent users. Today, this ratio has completely reversed—agent users now outnumber Tab users by a factor of two, and usage continues to surge. It was revealed that many Cursor users have now completely stopped using the Tab key.

The limitations of synchronous agents lie in a dual binding: they require real-time interaction with the developer while also competing for computational resources on the local machine. This severely limits the number of synchronous agents that can run concurrently. Cloud agents fundamentally remove these constraints. Each agent runs in an isolated cloud virtual machine. After a developer delegates a task, they can move on to other work without needing to monitor the process in real-time. Over spans of hours, the agent autonomously completes code iteration and testing, ultimately delivering results in the form of logs, recordings, or live previews, rather than displaying code changes line-by-line. This delivery method makes running multiple agents in parallel a reality. Developers can quickly assess the output quality of multiple tasks without needing to rebuild the context for each session from scratch. Consequently, the human role transforms fundamentally: from someone who "guides code line-by-line" to someone who "defines problems and sets review criteria."

Using its own internal practices as an example, Cursor outlines the specific shape of this new working model. Developers adopting the new approach exhibit three common characteristics: agents are responsible for writing nearly 100% of the code; developers concentrate their time on decomposing problems, reviewing artifacts and code, and providing feedback; and they initiate multiple agents to work in parallel, rather than guiding each one manually to completion. It was also acknowledged that large-scale industry adoption of this model still faces challenges. At an industrial scale, unstable tests or corrupted environments that an individual developer might work around can become systemic failures that interrupt every agent run. Furthermore, ensuring agents have access to the complete set of necessary tools and context remains a critical issue to be solved. Concurrently, it was stated that Cursor's recent feature releases represent an "initial but important step" in this direction.