Cursor launches agent mode, going head-to-head with Claude Code and Codex

Cursor, one of the fastest-growing startups in the history of AI developer tools, has announced the next generation of its product. It is no longer just an editor with smart autocomplete: Cursor is moving to an agent mode that can solve a developer's tasks on its own, from the initial request to the finished result. At the same time, the startup finds itself in direct competition with the world's two largest AI labs, Anthropic and OpenAI.

Anysphere, the company behind Cursor, followed a path typical of the best AI startups: it took a well-known tool (VS Code), added deep integration with language models, and built a product that developers came to love for its speed and convenience. By early 2025, Cursor had become one of the most talked-about IDEs among professional programmers, and the company's ARR surpassed $100 million, a record pace in the category's history.

Until now, Cursor was built on models from third-party providers: Claude from Anthropic, GPT-4 from OpenAI and others. The startup was effectively a reseller of compute with a better interface: deep codebase context, a smart chat, built-in refactoring commands. But with the launch of agent mode, Cursor takes a step toward autonomous task execution, and inevitably runs head-on into the same companies whose APIs it uses.

What does agent mode look like in practice? Previously, a developer gave Cursor a specific job: "write a function" or "fix this bug." Now the tool can accept a high-level task, for example "add OAuth2 support to this project," and work through the whole process itself: study the code structure, identify the relevant files, make the edits, run tests and propose a final result. This is a fundamentally different level of autonomy.

Competition in this space has already taken shape. Anthropic's Claude Code is a terminal agent that works with the file system, runs commands and iterates on a task. Its main strength is deep integration with Claude and support for the MCP protocol. OpenAI's Codex is a cloud agent that performs tasks asynchronously in isolated containers and can work on several tasks in parallel. Both products were released or substantially updated in 2025, and both target the same audience: developers who want not a suggestion assistant but a full-fledged AI colleague.

For Cursor, this is a strategic fork in the road. Agent mode is a logical evolution of a product developers already like. But Cursor depends on models from Anthropic and OpenAI. By launching an agent, the startup changes its position in the ecosystem and risks straining relations with suppliers that are now direct competitors.

The AI development tools market is consolidating quickly. A few months ago there was room for dozens of players: GitHub Copilot, Cursor, Windsurf, Replit, Devin, Codeium. Today the largest labs are building their own agent products and are competing no longer over a single feature, but over the right to become the infrastructure layer in a programmer's daily work.

For the average developer, this means one thing: in the coming months, the quality of AI agents will rise sharply, and competition will force every side to improve its tools quickly. Who will win this race is not yet clear. But developers stand only to gain from the fight.