跳到主要内容

Agents

AI that takes action, not just responds.

An agent is an AI system that can do things — read files, edit documents, run commands — rather than just answer questions.


Chat vs. Agent

Chat AIAgent AI
Answers questionsPerforms tasks
Output: text in a boxOutput: changes to your files
You copy-paste resultsAI applies changes directly
Stateless (each message is new)Contextual (knows your project)

When you use Claude Code, you're working with an agent. It doesn't just tell you what to do — it does it.


What Agents Can Do

Modern AI agents can:

  • Read your files and understand context
  • Edit documents directly
  • Create new files
  • Run terminal commands
  • Search your codebase
  • Ask clarifying questions

All with your approval before each action.


The Approval Loop

Agents don't act without permission. The typical flow:

  1. You give an instruction
  2. Agent analyzes what's needed
  3. Agent proposes an action
  4. You approve or reject
  5. Agent executes (or adjusts based on feedback)
  6. Repeat

This keeps you in control while letting AI do the heavy lifting.


Example Session

You: Check the grammar in chapter-1.md

Claude: I'll read chapter-1.md and check for grammar errors.
[Reads file]

I found 3 issues:
- Line 12: "industriel" → "industrial"
- Line 15: "transportaton" → "transportation"
- Line 23: "emphsized" → "emphasized"

Would you like me to fix these?

You: Yes

Claude: [Edits file]
Done. I've corrected all 3 spelling errors.

The agent read your file, identified issues, proposed changes, and executed them — all in one flow.


Why This Matters

The agent paradigm changes your relationship with AI:

BeforeAfter
You learn AI's limitationsAI learns your project
You adapt to AI's formatAI adapts to your workflow
You do the integration workAI handles the mechanics

You focus on what you want. The agent figures out how.


Available Agent Tools

ToolTypeDescription
Claude CodeTerminalAnthropic's CLI agent
CursorEditorAI-integrated IDE
WindsurfEditorAI-first code editor
AiderTerminalOpen-source alternative

They all implement the same core idea: AI that acts, not just answers.


Next: Learn the AGENTS.md Pattern — how to give agents persistent instructions.