Can switch back and forth between fast and slow models to speed up the workflow.
Plan mode is more reliable, build up the reasoning before you unleash agent mode on your codebase. Composer 1.5 is better for executing the plan. Heavier models are mainly good for reasoning, but aren't that helpful for the execution.
Agents can be predefined, added to the .agents folder.
Can dedicate seperate agents to specific worktrees so they can all work on the same repo together. Worktrees are in memory, and you can choose when to apply them. Cursor can help you with conflicts.
https://cursor.com/agents https://cursor.com/automations Long running process (hours, days) to tackle really long tasks Grind mode, (maximum effort) Can be used to create demos Automations - can be used to program forever tasks like ensuring good test coverage, fixing ci failures. Discovering CVEs.
Composer 1.5 Gemini 3 Flash (spend effecient)
Great for ask mode. Fast inference. Great for Agent mode after plan mode.
GPT-5.3 Codex Claude 4.6 Opus
Great for code reviews Great for plan mode