Cursor Agentv0.45

What changed when Cursor added a real agent

For years, Cursor was "VS Code with great inline AI." That's still true — but in late 2024 they shipped a background agent mode that elevates it from copilot to junior teammate. You hand it a task ("Add a /webhooks endpoint that verifies HMAC, queue jobs, write tests"), it disappears for 5-15 minutes, and comes back with a diff you can scroll through and accept hunk-by-hunk.

This is meaningful. Previously you had to drive Cursor with chat. Now Cursor drives itself. The difference shows up most on tasks where the spec is clear but the mechanical work is tedious — exactly the work that burns senior engineer time.

The diff UX deserves its own paragraph

Every coding agent shows you "here's what I changed." Cursor's implementation is the cleanest:

• File-by-file navigation, with a sticky overview panel.
• Hunk-level accept/reject, not just "approve all".
• Live preview of the resulting file as you toggle hunks on/off.
• Comments inline explaining why the agent made each change.

After a week of use, this UX becomes a moat. Switching to a tool that dumps a unified diff feels primitive.

Where it gets uneven

The single biggest performance variable is context selection. If you mention @codebase and let Cursor index everything, you get a 70%-good answer. If you carefully @-mention the 3 files that matter, you get a 95%-good answer. The skill ceiling on prompting matters more than people admit.

Two specific failure modes we hit repeatedly:

1. Large monorepos — past ~500k LOC, the indexer slows down and starts giving stale recommendations. Mitigate with .cursorignore listing build outputs, generated SDKs, and historical archives.

2. Context drift on long sessions — after ~30 minutes of conversation, Cursor sometimes forgets earlier decisions. Restarting the chat with a fresh prompt summarizing the goal usually fixes it.

Pricing math

Cursor Pro is $20/month — same as ChatGPT Plus or Claude Pro. For that you get:

• 500 "fast" requests (low-latency, frontier models)
• Unlimited "slow" requests (queue when fast is exhausted)
• Background agent access
• Repository indexing

For a developer who'd otherwise pay individually for Claude Pro + GitHub Copilot ($30/mo combined), Cursor Pro is a clear win. For a team, the Business plan is $40/seat/month and includes admin controls, privacy mode (your code never trains models), and SSO.

Honest comparison

• vs Cline: Cursor is more polished, less transparent. Cline is more transparent, less polished. If you want to see every prompt, choose Cline. If you want fewer decisions, choose Cursor.
• vs Devin: Devin runs longer and more autonomously, but costs 25× more. Cursor's background agent is the right tool for 90% of tasks Devin can do.
• vs Copilot: Different categories now. Copilot is autocomplete-plus. Cursor is closer to a junior dev sitting at your desk.

Bottom line

Cursor is the default recommendation for most developers in 2025. It's not the most autonomous, the most transparent, or the most experimental — but it's the agent that consistently delivers value without ceremony. If you're starting from zero, start here.