Cursor Rolls Out Big AI Upgrade As Coding Battle Heats Up

As the business fights off fiercer competition from rivals like Anthropic, OpenAI, and Microsoft, Cursor released improvements to its artificial intelligence coding agents on Tuesday.

Anysphere’s Cursor announced a significant increase in its AI agent capabilities, introducing Cloud Agents to turn the tool from a “copilot” to an independent “colleague”.

The AI startup told members of the press that its enhanced agents extend its capabilities to the next level and that it must continue to innovate to gain market share and new consumers. The startup reported in November that its annualised revenue has surpassed $1 billion, and its valuation has skyrocketed to $29.3 billion. Although Cursor entered the AI coding market somewhat early, other companies have already introduced rival products.

Artificial intelligence (AI) agents are tools that can do tasks for a user. As model capabilities have increased over the past year, their popularity has skyrocketed. Software engineers have been among the first to use agents from firms like Cursor to create, modify, and review code.

According to the business, Cursor’s enhanced agents are able to test their own modifications and document their work using screenshots, logs, and videos. Users can activate the agents via the web, Microsoft’s GitHub, Cursor’s desktop software, a mobile device, or the messaging app Slack.

The agents can also operate on their own virtual machines, which are cloud-based computers that function exactly like real ones, concurrently with complete development environments. According to Cursor, this eliminates the need for agents to waste time onboarding or fight for local resources on a developer’s laptop.

“You can have 10 or 20 of these things running at the same time instead of just one or three things,” Alexi Robbins, co-head of engineering for asynchronous agents Cursor, told CNBC in an interview. “This allows for a very high throughput.”

By February, OpenAI’s Codex had achieved 1.5 million weekly active users, while Anthropic’s Claude Code had increased to over $2.5 billion in run-rate revenue. In October, Microsoft CEO Satya Nadella reported that GitHub Copilot, the company’s tool, had over 26 million users. 

Established in 2022, Cursor gained a cult following the following year when it released the first iteration of its AI coding software.

Cursor’s upgraded agents can be given more difficult jobs by developers, and they can test and refine those features until they are finished. According to Cursor, this will free up developers’ time to concentrate more on matters of judgement and taste rather than modifying files and deciphering code.

“This is more like, ‘This is what it’s going to look like working with agents,’ rather than a new feature,” Jonas Nelle, Cursor’s other co-head of engineering for asynchronous agents, stated. “They’re kind of becoming full software developers; they’re not just writing code and software.

Nelle stated that Cursor has already experienced a “big transformation” as a result of testing the agents inside.

According to Cursor, agents running on their own virtual machines currently create about 35% of pull requests, which are requests from developers to make changes to the codebase.

“Working with these agents allows you to accomplish so much more as an individual,” Nelle stated. “That progression has undergone another step change.”

The update adds a number of crucial capabilities intended to autonomously manage intricate, multi-step engineering tasks, which are 

• Isolated Virtual Machines: Instead of vying for local resources on a developer’s computer, agents now operate on specialised, isolated cloud virtual machines.
• Parallel Execution: By triggering 10 to 20 agents at once, users can complete parallelisable tasks, such as refactoring several components or simultaneously addressing many issues, with enormous throughput.
• Self-Testing & Verification: Before a human examines their modifications, the agents can construct the software they create, conduct tests, and capture video proof (along with logs and screenshots) to demonstrate that their modifications are effective.
• Increased Access: Developers can assign tasks using natural language without being restricted by the IDE by calling these agents through Slack, GitHub, the web, and mobile devices.
• Long-running Research Preview: To plan and carry out more complex, long-term architectural activities, a new “Long-running Agent” mode is currently in preview.

This action comes after a period of fierce competition during which competitors like Anthropic’s Claude Code and GitHub Copilot (with its agent mode) have reduced the difference in autonomous coding features. A move toward “self-driving codebases” is indicated by Cursor’s current assertion that 30% of its own pull requests are now produced by its internal agents.