Agent overview
Three components#
Every agent is built from the same three pieces.
Instructions are the system prompt plus whatever rules you have defined. This is the layer you control most directly.
Tools are the actions available to it: reading and editing files, searching the repository, running shell commands, fetching from the web, and more. See tools for the full set.
Model is the frontier model you select for the task.
Cursor tunes instructions and tool definitions for each supported model rather than sending one generic prompt everywhere. As new models ship, that adaptation happens on Cursor's side.
What a task looks like#
Given a request, the agent searches for relevant files, reads them, proposes edits, and runs commands to check its work. There is no fixed limit on how many tool calls a single task can make, so a broad request may involve dozens of reads before the first edit appears.
This is why a specific request outperforms a vague one. "Fix the login bug" sends the agent hunting. "Users with expired sessions get a 500 instead of a redirect on /account" points it at the right file almost immediately.
Checkpoints#
The agent snapshots your files before significant changes, so you can preview and roll back if it takes a wrong turn. See checkpoints.
Queued messages#
You do not have to wait for the agent to finish before typing your next instruction.
Press Enter while it is working to add a message to the queue. Queued messages appear in order below the active task and run sequentially once the current one finishes. You can drag them to reorder.
Press Cmd+Enter to bypass the queue and send immediately. The message attaches to the most recent turn and is processed right away, which is what you want when the agent is heading somewhere wrong and you need to redirect it now.