The one-line difference
A copilot suggests while you type; an agent acts while you watch. The copilot’s unit of work is the suggestion inside your editor session. The agent’s unit of work is the task — it plans, runs tools, and finishes something, as in the agent definition.
Side by side
| Aspect | Coding copilot | AI agent |
|---|---|---|
| Primary loop | Autocomplete and chat in the editor | Plan → tool calls → verify → result |
| Tool access | Mostly editor actions | Shell, files, browser, APIs — under policy |
| Where it runs | Inside the IDE | Terminal, background service, cron |
| Who finishes the work | You write, it assists | It executes; you approve or review |
| Typical cost model | Subscription | Pay per token, controlled by budgets |
When to use which
- Use a copilot when your hands are on the keyboard and the editor context is the point.
- Use an agent when the work is a whole task: a repo to scaffold, a build to run, a demo like the one-prompt maps clone.
- They compose: let the agent do the multi-step build, then use the copilot for the fine edits.
The control difference
Because an agent can touch the shell and files, it needs what a copilot usually skips: secure mode, an audit trail, and sandboxing. Compare further with the full comparison.