Cursor 3: the AI-native IDE with parallel agents on git worktrees
Cursor 3 crossed $500M ARR and became the reference AI IDE among developers. It's a VS Code fork designed around AI: ultra-fast tab completion, agent mode, composer for multi-file refactors. The v3 headline: parallel agents. Up to 8 agents can run concurrently on different parts of your codebase via git worktrees, avoiding conflicts and parallelizing independent tasks (e.g., one agent on tests, one on docs, one on a back-end refactor). Cursor remains the go-to tool for developers who want to stay in a classic IDE while leveraging the latest models (Claude Sonnet, GPT, Gemini) with fine-grained orchestration.
For who
Front-end/back-end developers and product teams who want an AI IDE with parallel agents and an easy VS Code migration path.
Input
Prompts en langage naturel, code existant dans l'IDE, references a des fichiers et symboles.
steps (6)
Install Cursor 3
infoDownload Cursor from cursor.com for macOS, Linux or Windows. Import of your VS Code config is offered on first launch.
Pick your models
infoConfigure LLMs (Claude Sonnet, GPT, Gemini) based on speed/quality/cost preferences.
Use tab completions
infoStart coding: Cursor continuously suggests code anticipating your intentions.
Switch to agent mode
infoFor broader tasks, open the agent panel and describe the goal in plain English (e.g., "add a REST endpoint to delete a user").
Launch parallel agents
infoIn the Cursor 3 panel, trigger multiple agents simultaneously on different parts of the codebase via git worktrees.
Review the diffs
infoConsolidate agent work by reviewing diffs, merging worktrees, and running tests.
Output
Diffs de code, PR, suggestions tab completion, refactors multi-fichiers executes via composer.
Example
Input
In the agent panel: "Migrate all our Express routes to Hono, add tests for each new handler, and update the OpenAPI doc."
Output
Cursor 3 launches 3 parallel agents: one on /api/users, one on /api/orders, one on /api/billing, each in its worktree. Each agent lands commits and 10+ tests. You merge the three worktrees and run CI.
Customization
| Parameter | Description | Default |
|---|---|---|
| Modele principal | LLM utilise par defaut par les agents (Claude Sonnet, GPT, Gemini). | Claude Sonnet |
| Nombre d'agents paralleles | Nombre maximum d'agents simultanes (jusqu'a 8). | 3 |
| Mode privacy | Activer le Privacy Mode pour ne pas envoyer le code aux fournisseurs LLM au-dela de l'inference. | Active |