Cursor 2.4、サブエージェント・Skills・画像生成を追加 Subagents, Skills, and Image Generation

Cursor has shipped version 2.4, bringing three notable additions to its AI-native IDE: subagents, Skills, and built-in image generation. The release continues the trend of pushing the editor beyond autocomplete into a full agent-driven development environment.

The subagent feature allows the main agent to spin off child agents that work in parallel on delegated tasks. This is particularly useful for activities like codebase exploration, isolated refactors, or running multiple investigations concurrently without polluting the primary agent's context window. By offloading work into separate contexts, users can maintain focus in the main thread while still progressing on auxiliary tasks. The pattern echoes Anthropic's subagent concept in Claude Code, suggesting that multi-agent orchestration is rapidly becoming a standard primitive across coding assistants.

Skills introduce a way to package reusable instructions, prompts, and tool-usage patterns into modular units that the agent can invoke on demand. Teams can encode project-specific coding conventions, deployment procedures, or testing strategies as Skills, reducing the need to re-state context each session and producing more consistent behavior across runs. The naming and structure closely mirror Anthropic's recently announced Agent Skills, and it appears Cursor is aligning with what may emerge as a de facto convention for portable agent capabilities.

Image generation rounds out the release. Developers can now generate UI mockups, icons, or documentation diagrams directly inside the editor, lowering the friction of moving between design tools and code. Cursor has not detailed which underlying model powers the feature, though it is likely routed through an external image API rather than running locally.

The broader context matters here. Cursor faces sharpening competition from GitHub Copilot's agent mode, Anthropic's Claude Code CLI, Windsurf, and a wave of open-source agent frameworks. Each of these is converging on a similar architecture: a primary planning agent, specialized subagents, reusable skill modules, and richer multimodal inputs and outputs. Cursor's 2.4 release reads as a deliberate move to keep parity on these axes while leveraging its strength as a polished editor experience.

For practitioners, the practical question is how subagents and Skills compose in real workflows. Parallelism is appealing, but it can introduce coordination overhead and conflicting edits if not carefully scoped. Skills, meanwhile, are only as valuable as the discipline teams put into authoring and maintaining them; poorly written Skills risk becoming stale prompt debt. Organizations adopting these features will likely need lightweight governance around what gets promoted into a shared Skill library, similar to how internal component libraries or lint rules are curated today. Overall, 2.4 is an incremental but strategically important step toward Cursor's vision of an agent-first IDE.