Subagents, Skills, and Image Generation

Source: Cursor Date: 2026-01-22 URL: https://cursor.com/changelog/2-4

Summary

Cursor 2.4 ships subagents (independent parallel agents for codebase research, terminal commands, and parallel work streams), a Skills system (agents discover domain-specific SKILL.md files at runtime for focused context), and direct image generation from text descriptions or reference uploads. Enterprise gets Cursor Blame for AI attribution tracking. Performance: in-editor browser 10x faster, hook commands 40x faster, MCP servers load dynamically. PDF reading added for agents.

Implications

Agent-IDE feature race. The Skills/SKILL.md system is a direct parallel to Claude Code’s CLAUDE.md — project-level instruction files that shape agent behavior. Cursor shipping this as “Skills” rather than “rules” signals a strategic framing difference: skills are composable and discoverable, rules are static config. Combined with subagents, Cursor is building a multi-agent coordination layer inside the IDE that CLI tools don’t have.

MCP/protocol adoption. Dynamic MCP server loading (reducing token usage) shows Cursor is treating MCP tool bloat as a real problem and solving it architecturally. This will benefit complex MCP setups and signals Cursor is invested in MCP as production infrastructure.

Watch: Whether SKILL.md files become a community ecosystem (shared skills packages) and whether Cursor Blame (AI attribution) becomes a compliance requirement that pushes enterprise competitors to ship similar features.