Cursor 3's Multi-Agent Pivot: Features vs High Costs
Cursor 3 shifts from IDE to multi-agent workspace for parallel coding tasks across models and repos, delivering working CRUD apps in 3-9 minutes, but burns $5 on simple tests—10x pricier than native tools.
Multi-Agent Workspace Enables Parallel Coding Experiments
Cursor 3 introduces a Conductor-like interface for running multiple agents in parallel across workspaces (local Mac, SSH, or cloud). Open separate folders, assign models like Composer 2, GPT-4o, or Opus, and issue identical prompts—e.g., "Add categories table to Laravel starter with posts relation, index page, tests." Agents generate code, run tests, commit to new branches (named with "cursor-"), push, and create GitHub PRs marked "Made with Cursor."
This workflow skips traditional IDE views by default (terminal and files accessible via top-right buttons), prioritizing agent supervision. Review diffs before committing; PRs include summaries, database changes, and verification steps. For simple CRUD, Composer 2 finishes in 3:21 (passing tests after manual npm run build), GPT-4o in 8:50 (card layout, includes post counts), and Opus slower (1/7 tasks at 5 minutes, needs manual migrations). All produce functional apps post-fixes, but require supervision for builds/migrations—Composer is fastest but shallower (misses post counts).
Switch agents mid-task or fall back to classic Cursor IDE, making it hybrid for prompt-review-commit cycles.
Model Performance Highlights Trade-offs in Speed vs Depth
Parallel testing reveals model quirks: Composer 2 excels in speed for day-to-day tasks (~1,000 fewer lines than GPT), but lacks depth (no post counts). GPT-4o and Opus add smarts like counts, with Opus generating comparable code despite slowness in Cursor (better in native Claude/Codex). Cursor may optimize Anthropic models, but native environments (Claude.dev for Opus, Codex for GPT) run faster/cheaper.
Commits use separate branches/PRs; no co-author like Claude. Multi-workspace supports model comparison without setup, ideal for benchmarking prompts across frontier models.
Cloud Agents Extend Access but Introduce Friction
Launch cloud agents on cursor.com/agents for remote execution (e.g., change "Get Started" to "Hello" in Laravel welcome.blade). Sets up Ubuntu VM with Cursor dependencies (36s), auto-installs PHP/Composer if missing via sub-agents (sleep/polling), writes tests, runs them post-fixes (e.g., php artisan key:generate), then commits/PRs (author: "cursor").
Succeeds after ~5+ minutes but fails initially on missing tools/exceptions. Max context mode inflates costs; VMs lack stack-specific setups (e.g., no PHP), forcing installs. Use for laptop-free scenarios, but local is faster/cheaper/reliable.
Pricing Burns Through Limits on Routine Tasks
$5 daily usage (from $20 Pro plan) for three local CRUD agents + one cloud text change—~8 similar runs exhaust monthly quota. Breakdown: Cloud Opus (700k tokens, ~$1), local Opus (2.8M tokens), Composer (~95¢). Subscriptions pass near-API costs without discounts.
Competitors differentiate: Conductor ($22M raised) integrates Claude/Codex subscriptions (gray area); SolarTerm/Poliscope launch native terminals. Cursor's model-agnostic middle-ground shines strategically for agentic futures (prompt-review-reprompt), but 10x cost vs. $20-25 Claude/Codex plans kills adoption for heavy use.