Cursor 3 ('Glass') shifts from AI-assisted coding to autonomous agent tasks, directly competing with Anthropic's Claude Code and OpenAI's Codex.
Cursor launched Cursor 3, codenamed 'Glass', an agent-first coding interface where developers assign tasks to AI agents rather than requesting inline suggestions. The product directly targets Claude Code and OpenAI Codex, which have rapidly gained developer adoption with heavily subsidized subscriptions. Cursor, which built its business on top of OpenAI and Anthropic's models, now competes head-to-head with those same labs for enterprise developers. The shift signals a fundamental rethink of Cursor's product: the IDE paradigm that made them successful is becoming less central.
Cursor 3 moves from autocomplete-and-assist to task-delegation: you describe work, an agent executes it. This puts it in direct technical competition with Claude Code's terminal-native agent and Codex's cloud-sandboxed approach. The architecture choice matters — IDE-embedded agents (Cursor) vs. CLI-first agents (Claude Code) vs. cloud-isolated agents (Codex) will produce meaningfully different DX for different codebases and team sizes.
Run a real multi-file task — e.g. 'refactor this auth module to use JWTs' — in both Cursor 3 and Claude Code this week, then compare: number of human interventions required, files touched correctly on first pass, and wall-clock time to a passing test suite.
Suggested exercise · not independently tested. Access requirements and costs may vary.
Download or update to Cursor 3 from cursor.com and open an existing project
In the agent panel, type: 'Refactor the user authentication flow to replace session tokens with JWTs. Update all affected routes, middleware, and tests.'
Watch the agent's step-by-step execution trace — note how many times it asks for confirmation vs. acts autonomously
Run your test suite and record pass rate on first attempt
Tags