Cursor 2.2: Visual Editor, Debug Mode & the Rise of AI Subagents

Cursor
, ,
1

Cursor has been on a tear lately. The 2.0 release in October 2025 was a big deal — introducing their own Composer coding model and a redesigned agent-centric interface. But the follow-up releases through late 2025 and into 2026 have added features that fundamentally change how you work with the editor.

Visual Editor: Design in Your IDE

The headline feature from Cursor 2.2 (December 2025) is the Visual Editor — and it’s exactly what it sounds like. You can now visually interact with your web application directly inside Cursor’s built-in browser:

  • Drag-and-drop: Rearrange UI elements directly on a live web app canvas
  • Component inspection: Click any element to see its React props in a sidebar
  • Visual property controls: Adjust styles with sliders, color pickers, and design tokens
  • Point-and-prompt: Click on any element and describe changes in natural language — agents apply them to the underlying code

This effectively merges what you’d do in Chrome DevTools with what you’d do in Figma, all inside your code editor. The practical impact is significant for frontend work: you see the change, click the element, describe what you want, and the code updates.

That said, some developers have noted that every visual change requires an AI agent call, which means each tweak costs credits. For small styling adjustments, that can feel heavy-handed. It’s a valid concern worth watching as Cursor refines the pricing model.

Debug Mode: How AI Should Debug

This is the feature that impressed me most. Instead of the typical AI approach of generating speculative fixes, Cursor’s Debug Mode takes a fundamentally different approach:

  1. Watches your app fail — observes the actual runtime behavior
  2. Auto-instruments with logging — adds targeted log statements to understand execution flow
  3. Reads real data — analyzes actual error logs and runtime data
  4. Delivers precise fixes — typically 2-3 lines instead of 40-50 lines of guesswork

This is how experienced developers debug. They don’t guess — they observe, instrument, analyze, then fix. Having an AI that follows this same process is a genuine step forward.

BugBot: Automated PR Review That Actually Works

BugBot has matured significantly since its July 2025 launch. It’s an automated PR code reviewer that catches issues before merge and includes “Fix in Cursor” prompts that jump you directly to problematic code.

The numbers tell the story: since Version 1, Cursor ran 40 major experiments to improve quality. The resolution rate went from 52% to over 70%, and the average bugs flagged per run went from 0.4 to 0.7. That means resolved bugs per PR more than doubled from roughly 0.2 to 0.5.

For teams, this is significant. It’s not replacing human review, but it’s catching real issues that humans miss — especially the subtle ones that slip through in large PRs.

Multi-Agent Judging

When multiple AI agents produce different solutions to the same problem, Cursor can now automatically evaluate and select the best one with a justification. This is particularly useful when you’re using Plan Mode and the agents take different approaches — instead of you comparing outputs, Cursor does the comparison and explains its reasoning.

Sub-Agents and Agent Harness Improvements

The latest releases (January-February 2026) brought substantial agent infrastructure improvements:

  • Subagents: Independent agents specialized for discrete parts of a parent task, running in parallel with their own context, custom prompts, and model selections
  • Custom slash commands: Reusable prompts saved in .cursor/commands/*.md files
  • Hooks (beta): Custom scripts that observe and control agent behavior at runtime
  • Team Rules: Global policies that apply across all projects, including BugBot rules
  • Cursor Blame (Enterprise): Tracks which code came from Tab completions, agent runs, or human edits — broken down by model

Cursor Blame: AI Attribution

This one is particularly relevant as AI-generated code becomes the norm. Cursor Blame distinguishes between code from Tab completions, agent runs (broken down by model), and human edits. It lets teams track AI usage patterns across their codebase.

For engineering managers and leads, this is valuable data. You can see which models produce the most reliable code, track how much of your codebase is AI-generated, and make informed decisions about model selection and review processes.

The Bigger Picture

Cursor has over 50% of Fortune 500 companies as users, including Nvidia, Uber, and Adobe. The $200/month Ultra plan launched in June 2025 signals their confidence in delivering ROI for power users.

The trajectory is clear: Cursor is evolving from “VS Code with AI” into an agent workbench that happens to be an editor. The Visual Editor, Debug Mode, and sub-agent system are steps toward a future where you describe what you want built and the editor orchestrates the work.

For those of you using Cursor — has the Visual Editor changed your frontend workflow? And how do you find the credit costs compared to what you’d spend on other tools? I’m particularly interested in hearing from anyone using it on a team with BugBot enabled.