"AI Programming Star" Cursor CEO: The "Third Era" of AI software development has arrived

AI-assisted programming is undergoing a fundamental paradigm shift.

Recently, Cursor co-founder Michael Truell posted on X platform stating 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 long periods. This means Cursor’s positioning has fundamentally shifted from a “code-writing tool” to a “platform that helps developers build software factories.”

Data confirms this judgment: among the code submissions (PRs) merged within Cursor, 35% are completed by autonomous intelligent agents running on cloud virtual machines. More notably, the company expects that within a year, the vast majority of software development work will be carried out by these types of agents. This trend will not only reshape the competitive landscape of AI programming tools but also have a profound impact on the business models of the entire industry.

From Tabs to Agents: Rapid User Behavior Reversal

The pace of AI programming evolution is exceeding all expectations. Truell reviews the key turning points in this field, clearly dividing the development of AI programming tools into three stages:

Stage One: The Tab Auto-Complete Era. Tabs can not only complete the current line but also intelligently predict and finish the next line of code, even making multiple modifications across files, freeing developers from trivial coding tasks. This stage lasted nearly two years, focusing on automating low-entropy, repetitive work.

Stage Two: The Synchronous Agent Era. The core feature of this era is conversational programming, where developers describe their needs to the agent using natural language, and the agent generates code in real-time and responds, forming a rapid “prompt-feedback-correction” interaction loop. He predicts that this stage may last less than a year, with the transition happening much faster than previously expected.

Stage Three: The Cloud Agent Era. After assigning tasks, developers let the agents run independently on cloud virtual machines—autonomously completing coding, debugging, testing, and iteration—shifting the developer’s role from “person writing code” to “person commanding the agent.”

User behavior data confirms the dramatic nature of this paradigm shift: in March 2025, Cursor’s Tab users outnumbered agent users by 2.5 times; now, that ratio has completely reversed—agent users are twice the Tab users, and usage continues to surge. He reveals that many Cursor users have now completely stopped using the Tab key.

Core Advantages of Cloud Agents: Parallelism and Asynchrony

The limitation of synchronous agents lies in dual binding: they require real-time interaction with developers and must compete for computing resources on local machines. This means the number of simultaneously running synchronous agents is very limited.

Cloud agents fundamentally eliminate these constraints. Each runs on an independent cloud virtual machine, so once a task is handed over, developers can move on to other matters without waiting in real-time. Over a span of hours, the agents autonomously complete code iterations, testing, and validation, ultimately delivering results via logs, recordings, or real-time previews, rather than line-by-line code diffs.

This delivery method makes it feasible to run multiple agents in parallel. Developers no longer need to rebuild context from scratch for each session, enabling quick evaluation of multiple task outputs. The human role thus undergoes a fundamental transformation: from “guiding code line-by-line” to “defining problems and setting review standards.”

Internal Practices Reveal New Developer Work Modes

Using internal practices as an example, Cursor describes the concrete form of this new working mode. Developers adopting this approach share three common features: agents are responsible for nearly 100% of code writing; developers focus on breaking down problems, reviewing artifacts and code, and providing feedback; and multiple agents are launched in parallel rather than guiding each one step-by-step to completion.

He also admits that this model still faces challenges for large-scale industry adoption. In industrial-scale environments, unstable tests or environment damages that a single developer might bypass could evolve into systemic failures that interrupt each agent run. Additionally, ensuring agents have access to all necessary tools and context remains a key unresolved issue.

Meanwhile, he states that Cursor’s recent new feature releases are a “preliminary but important step” toward this direction.

Risk Warning and Disclaimer

Market risks exist; investments should be cautious. This article does not constitute personal investment advice and does not consider individual users’ specific investment goals, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions herein are suitable for their particular circumstances. Invest at your own risk.