The copy-paste tax
If you've coded with an AI chat window, you know the routine: paste a snippet, get a snippet, paste it into your editor, run it, paste the error back, repeat. You are the integration layer, and that's slow and error-prone.
What changes with an agent
| AI chat | Claude Code | |
|---|---|---|
| Context | What you paste | Your whole project, searched as needed |
| Changes | Suggestions | Real edits, with your approval |
| Feedback loop | You run things | It runs tests and builds itself |
| Memory | Per chat | CLAUDE.md, skills, settings in the repo |
| Automation | None | Headless runs, CI, schedules |
The big unlock: verification
Because Claude Code can run your tests, you can say "fix it and prove it." The agent iterates until the check passes, instead of handing you code that looks right. This one shift (from plausible to verified) is why agentic coding feels so different.
When chat is still fine
Quick conceptual questions, explaining a snippet or brainstorming names: chat is great. The moment the task touches real files, needs running, or spans more than one place, an agent wins.
Try the difference
Spend 9 minutes on Lesson 1 and watch a simulated session fix a real bug end to end.