Cursor ditches VS Code, but not everyone is happy...

High-level summary

The Code Report (video, 2026-04-06) analyzes Cursor 3 — a major pivot from a VS Code fork to an agent-orchestration IDE — and the new in-house model Composer 2. The video demos building a prototype using swarms of coding agents and covers features, performance claims, controversies, and community reaction.

Key product / tech changes in Cursor 3

• Rewritten architecture

  • No longer a VS Code fork: rebuilt from scratch in Rust (for performance/memory benefits) and TypeScript.
  • The legacy VS Code-style editor remains available but is de-emphasized.

• New UX: agent-centric workspace

  • Workspace acts as a home for agents and combines project files, language servers, git history, terminal, and a minimal file explorer.
  • Remote SSH support and the ability to run agents across multiple repositories, machines, and cloud instances.
  • Built-in browser to preview and interact with running apps directly inside the IDE.
  • Design mode: highlight UI elements and instruct the AI to change CSS/UI; changes are applied in the background so you can keep queuing tasks.

• Agent swarm orchestration

  • “Plan mode” to generate architecture and execution plans.
  • Run many agents in parallel across projects and machines.
  • Agent status indicators: yellow = needs human permission for risky/unsafe commands; blue = completed and ready for review.
  • Fine-grained monitoring and the ability to review code and commits produced by agents.

• Positioning

  • Cursor 3 aims to shift development workflows from hand-coding toward managing autonomous agents that can generate large codebases quickly.

Composer 2 model (Cursor’s in-house model)

• Marketing claims

  • Composer 2 is marketed as faster, cheaper, and smarter than Claude Opus 4.6 based on Cursor’s internal benchmarks.

• Transparency controversy

  • Composer 2 was later found (via model ID in metadata) to be based on Moonshot’s Kimmy K2 model.
  • Cursor apologized and published a technical report describing their reinforcement learning and other modifications.

• Practical implication

  • Whether truly in-house or fine-tuned from Kimmy/K2, a performant, low-cost coding model is valuable for enabling a “zero code / agent-led” development workflow.
  • The provenance debate raises trust questions, but model performance and cost remain central to adoption.

Demo / tutorial highlights

• Real-world demo: prototype called “Horse Tinder”

  • Starts from a fresh project and uses plan mode to organize work.
  • Spins up parallel agents for the core app, landing page, and remote cloud work.
  • Agents generate roughly 13,000 lines of code within minutes.
  • Uses the built-in browser to view the app (includes an SVG Clydesdale), edits UI via design mode, and queues multiple fixes while agents run in the background.
  • Shows git history, terminal inspection, and remote SSH usage.

• Emphasis on speed

  • Host contrasts timelines: manual coding would take months, single-developer “vibe coding” would take weeks, while an agent swarm can deliver in minutes or hours.

Criticism and community reaction

• Product direction

  • Some users are unhappy with Cursor shifting from a traditional code editor to an agent orchestration platform.

• Concept overlap and comparisons

  • Complaints about similarity to OpenAI’s agent/codec concepts.

• Trust and provenance concerns

  • Trust issues arose after Composer 2’s provenance revelation (its Kimmy K2 origin and previous accusations that Kimmy was trained on Claude outputs).
  • Community debate centers on transparency versus practical utility: is it acceptable to deliver faster/cheaper results if model provenance is unclear?

Sponsor mention

Blacksmith: promoted as a drop-in replacement for GitHub Actions runners. Claimed benefits: - Twice as fast and 75% cheaper by running on bare-metal gaming CPUs (optimized for high single-core performance). - Provides CI observability (useful when agents produce lots of code). - Promotional offer mentioned: 3,000 free minutes/month.

Takeaway / implications

• Cursor 3 marks a strategic shift from an AI-assisted code editor to an environment for running and supervising autonomous coding agents at scale.
• If agent models (Composer 2 / Kimmy K2 variants) are truly faster, cheaper, and reliable, they could accelerate agent-led development workflows.
• Key concerns remain around transparency, provenance, and safety — Cursor includes permission prompts for risky commands, but trust will influence adoption.

Main speakers / sources mentioned

• The Code Report (video host / narrator)
• Cursor (company) — Cursor 1.0, 2.0, and 3.0 products and Composer 2 model
• Composer 2 (model) / Moonshot’s Kimmy K2 (model origin/provenance)
• Claude Opus 4.6 (Anthropic; benchmark comparison)
• OpenAI codecs (comparison / community criticism)
• Blacksmith (sponsor; GitHub Actions runner alternative)