Subagents, Skills, and Image Generation
CursorCursor released version 2.4 introducing three major capabilities: subagents that run in parallel with isolated contexts and can be configured with custom prompts and models, skills defined in markdown files that provide domain-specific knowledge and workflows discoverable by agents, and native image generation powered by Google Nano Banana Pro. The release also included Cursor Blame for Enterprise customers, which extends git blame with AI attribution tracking.
Sources & Mentions
2 external resources covering this update
Major Features
Subagents
Cursor introduced subagents as specialized independent agents designed to handle discrete components of a parent agent’s task. These subagents operate in parallel rather than sequentially, allowing multiple aspects of complex work to progress simultaneously. Each subagent maintains its own isolated context, preventing interference between different workstreams.
Developers can configure subagents with custom prompts tailored to specific tasks, control which tools each subagent can access, and select different models for different types of work. Cursor provides default subagents optimized for common workflows, including a research subagent for exploring codebases and understanding existing implementations, a terminal subagent for executing commands and managing development environments, and parallel work stream subagents for handling multiple independent tasks simultaneously.
Skills
Cursor introduced Skills as a new way to provide agents with domain-specific knowledge and procedural expertise. Skills are defined in SKILL.md files that contain custom commands, specialized workflows, and instructions for handling particular types of tasks.
Agents automatically discover relevant skills when working on tasks that align with the skill’s domain. Developers can also explicitly invoke skills using the slash command menu. This dynamic discovery mechanism provides significant advantages over static rules, as skills can include contextual information that adapts to the specific task at hand.
Image Generation
Cursor added native image generation capabilities directly within the agent interface. Developers can describe images in natural language, and the agent will generate appropriate visuals using Google’s Nano Banana Pro model. Generated images appear as inline previews within the agent conversation, and Cursor automatically saves generated images to the project’s assets/ folder by default.
Enterprise Features
Cursor Blame
Cursor introduced Cursor Blame as an Enterprise-tier feature that extends traditional git blame with AI attribution capabilities. While standard git blame shows which developer made each change, Cursor Blame distinguishes between changes made via Tab completions, changes generated by agent runs (tracked by model), and direct human edits.
Interaction Improvements
Agent Clarification Questions
Cursor enhanced agent interactions by allowing agents to ask clarifying questions mid-conversation without pausing their work. When an agent needs additional information, it can formulate questions while continuing to read files, make edits, and execute commands.
Platform Improvements
The release included 29 improvements across the platform, including bringing CLI agent mode support to the command-line interface, enabling dynamic loading of MCP servers via JSON configuration files, allowing mode switching mid-conversation, adding support for PDF attachments in agent conversations, and introducing service account linking for CLI usage. Performance improvements included hooks executing 40 times faster at startup and browser navigation operating 10 times faster within the editor.