Cursor 2.4: Subagents, Skills, and What They Mean for Your Workflow

What's New in Cursor 2.4

Released January 22, 2026, Cursor 2.4 focuses on agent infrastructure. The three major additions:

1. Subagents — Independent agents that handle discrete subtasks in parallel
2. Skills — Reusable domain knowledge defined in SKILL.md files
3. Image generation — Create mockups and diagrams directly in the editor

Subagents: Parallel Specialized Execution

Subagents are child agents that run independently of your main conversation. Each gets its own context window, tool access, and can use a different model.

Cursor includes default subagents for:

• Researching your codebase
• Running terminal commands
• Executing parallel work streams

You can also define custom subagents for your team's specific needs.

How This Fits Our Methodology

In our 7-step process, subagents expand what you can delegate:

Step	Before 2.4	With Subagents
Plan	AI reads specs, you analyze	Subagent maps dependencies while you focus on priorities
Build	Sequential code generation	Parallel generation across files
Test	One test file at a time	Multiple test suites simultaneously

The key shift: you're now delegating coordination, not just individual tasks.

Skills: Encoding Team Knowledge

Skills are defined in SKILL.md files. Unlike rules (which are always-on), skills are invoked when relevant.

Use skills for:

• Team-specific patterns ("how we build API routes")
• Complex workflows ("our deployment process")
• Domain knowledge ("our data model conventions")

Example: API Route Skill

# API Route Skill

When creating API routes in this project:

1. Use the handler pattern from `src/lib/api-handler.ts`
2. Add Zod validation with schemas from `src/types/`
3. Include rate limiting via the `withRateLimit` wrapper
4. Log to our standard format

The agent discovers and applies this when you ask it to create an API route.

Image Generation

Generate images directly from agent conversations. Useful for:

• UI mockups during planning
• Architecture diagrams
• Quick visual assets

Images save to assets/ by default.

Cursor Blame (Enterprise)

Cursor Blame adds AI attribution to git blame. For each line, you can see:

• Whether it came from Tab, Agent, or human
• Which model generated it
• A link to the conversation that produced it

This matters for code review. When reviewing AI-generated code, you want the context of why it was written that way.

Practical Application

What to Delegate Now

With 2.4, you can confidently delegate:

• Multi-file scaffolding (subagents handle coordination)
• Pattern-following tasks (skills encode your patterns)
• Initial test coverage (parallel test generation)

What Still Needs Human Review

• Architecture decisions (subagents execute, they don't decide)
• Security-critical code (AI doesn't understand your threat model)
• Business logic (domain knowledge beyond your skills files)

What Humans Own

• Strategic direction
• Final approval
• Production decisions

This hasn't changed. The tools are better, but the judgment stays with you.

---

We train teams to use these features effectively. Contact us for hands-on training with Cursor 2.4.