Cursor 3.0: Parallel Agents and the /best-of-n Command

Cursor 3.0 dropped on April 6, 2026, and it's a bigger shift than a typical version bump. The product is no longer positioned as an AI-enhanced code editor — it's now an agent orchestration workspace. The headline feature is the new Agents Window, which lets you run multiple AI agents simultaneously across different repos, environments, and even remote machines.

Whether this matters to you right now depends on how you work. Solo developers on single projects will notice a faster, more capable editor. Teams running multiple services in parallel, or anyone who wants to compare AI outputs across models, will find the new features genuinely useful.

The Agents Window: Running Tasks in Parallel

The Agents Window is the centerpiece of Cursor 3.0. It lets you spawn multiple agents and have each one work on a different task at the same time — across different repos, local folders, cloud environments, or remote SSH sessions. You can watch them work, switch between them, and pull results from whichever finishes first.

Practical use cases: run one agent writing unit tests while another writes documentation. Run one agent refactoring a module while another checks for breaking changes in dependent code. Or use it to parallelize a migration — split the work across agents by subdirectory and have them run simultaneously.

How to use the Agents Window:
1. Open Cursor 3.0 and press Cmd+Shift+A (or find Agents in the View menu)
2. Click 'New Agent' to spawn a fresh agent with its own context
3. Each agent gets its own conversation, file access, and terminal
4. Agents run independently — one can be building while another is testing
5. Switch between agents with the tabs at the top of the Agents Window

The /best-of-n Command: Pick the Best Output

This is the feature most developers will use immediately. Type /best-of-n before any task, and Cursor runs that task simultaneously across multiple models. You see all outputs side by side and pick the one you want to apply.

It's useful when you're not sure which model handles a specific task best, or when a task is important enough that you want to compare a few approaches before committing. Think of it as peer review from multiple AI models at once.

Example usage:

/best-of-n Write a function that parses a CSV with potentially malformed rows, handles encoding issues, and returns a list of dicts with cleaned values. Add error logging.

Cursor runs this across 3+ models simultaneously. You'll see each implementation, review the differences, and accept the one that looks most correct or most aligned with your codebase style.

Self-Hosted Cloud Agents for Enterprise

Cursor 3.0 also added self-hosted cloud agents for enterprise customers. Instead of code execution happening on Cursor's servers, your build outputs, secrets, and file access stay inside your own infrastructure. This was the primary blocker for many companies adopting Cursor — proprietary code leaving their environment.

If you're at an enterprise evaluating AI coding tools, self-hosted agents are now on the table for Cursor. It's worth revisiting any security reviews that previously ruled it out.

What's the Same

The core editing experience — Composer, Chat, Tab autocomplete, .cursorrules, and keyboard shortcuts — is unchanged. If you've been using Cursor effectively before 3.0, you don't need to relearn anything. The new features layer on top of what you already know.

Should You Update Now?

Yes. Cursor 3.0 is a free update. The Agents Window and /best-of-n are opt-in features — you don't have to use them if your current workflow is working. But /best-of-n alone is worth the update if you ever find yourself second-guessing which model would write better code for a specific task.