Cursor Ask vs Agent vs Composer: The Mental Model

Ask, Agent, Plan and Debug are modes. They all run on the model you picked. Switching mode swaps the toolset the agent can reach for; the underlying model stays put. Pick the mode by how much you want the agent to do.

Read the diagram as two columns you set independently. Left is the mode dropdown, which decides what the agent is allowed to touch. Right is the model picker: it decides which model reasons about the problem and writes the code. Set each one separately; they don't move together. The table below spells out what each mode can and can't do.

One part of that field note has since moved on. As of 2026-07-22, Auto on Teams and Enterprise plans is powered by Cursor Router: a classifier runs on each agent request and routes it by task type and complexity, and what you select is an optimization mode - Cost, Balance or Intelligence - not a model. Cost keeps the previous Auto routing logic and its bundled Auto pricing; Balance and Intelligence bill at the routed model's rate. The router is on by default for Teams plans, and Enterprise admins turn it on from the dashboard.

Pick the mode by how much you want the agent to change and how risky that change is. Walk down this list and stop at the first line that matches your task.

1. 1Just need to understand something? → Ask mode (read-only, no diffs to review).
2. 2One small, known change in a file you have open? → Inline edit (⌘/Ctrl+K).
3. 3Chasing a bug or a failing test? → Debug mode.
4. 4A change that spans files or needs commands run? → Agent mode.
5. 5Big or risky task? → Plan mode first: it researches and writes an editable plan with no code changes, you review it, then switch to Agent to execute.

Whole disciplines live mostly in one mode. If your day is research and understanding rather than shipping edits (support, onboarding onto an unfamiliar codebase, answering a security questionnaire against your own repo), Ask is the efficient default. It gives you semantic search and codebase understanding with no write, edit or terminal access, so you get the answer without burning tokens on a full agent run or risking an unwanted change. Pair it with a cost-effective model and it lands the same answer Agent would.

Since Cursor 3Cursor's agent-forward interface (also called the agent window or Glass), built to run and supervise many agents at once rather than edit one file. Press Enter for the full definition..11 you also don't have to choose between asking a question and letting a long agent run finish: side chats open a parallel thread for questions and tangents without interrupting the main agent conversation.

Mode and model are separate dials, so picking the mode leaves the model open. A clean way to choose: weigh every model on three axes (performance, cost and speed) and accept that different tasks want a different mix. Planning leans on performance. Execution leans on speed. Most teams are cost-insensitive on the thinking step and cost-sensitive on the grind.

That three-axis view collapses into one habit that shows up in every workshop: plan with the smart model, build with the fast one. Use a high-reasoning frontier model (GPT-5.x, Opus 4.x) to research and write the plan, then switch to ComposerCursor's own fast coding model, tuned for the editor and priced well below frontier models; the recommended day-to-day model for executing a plan. Press Enter for the full definition. to execute it. By the time the plan exists, the hard thinking is done; what you want from execution is speed, and Composer has enough context and juice to carry it out.

The split also covers non-code work. ComposerCursor's own fast coding model, tuned for the editor and priced well below frontier models; the recommended day-to-day model for executing a plan. Press Enter for the full definition. is engineering-tuned, so it shines on data and analytical tasks like queries and reconciliations. For prose (an accounting memo, a write-up), a general-purpose model fits better. If you'd rather not pick per task, run Auto: on Teams and Enterprise plans the router classifies each request by task type and complexity and picks the model for you.

The usual mix-up is treating mode and model as the same dial. They aren't. Mode changes which tools the agent can touch; the model picker changes which AI runs. The ComposerCursor's own fast coding model, tuned for the editor and priced well below frontier models; the recommended day-to-day model for executing a plan. Press Enter for the full definition. name makes it worse, because it sounds like a mode but is actually a model. And most beginners just default to Agent, even for questions where Ask is safer and faster. The fix is the decision model above: pick the mode by the task's blast radiusHow much breaks if a change goes wrong; the scope of potential damage. Press Enter for the full definition., and treat the model picker as a separate choice.