Why This Works: Agent Platform vs Consumer App
Outcome: You understand the paradigm shift and key features of agent platforms
The Key Insight
What you just did in section 1.2? You couldn't do that in browser Claude.
If you went to claude.ai right now and tried the same thing, it wouldn't work. It can't write files. It can't access your documents. It can't create a to-do list for itself.
Why? Because browser Claude is a consumer app. Claude Code is an agent platform.
Understanding this difference is the foundation of everything else in this course.
Consumer App vs Agent Platform
| Feature | Consumer App (Browser) | Agent Platform (Claude Code) |
|---|---|---|
| Chat with AI | Yes | Yes |
| Access your files | No | Yes |
| Create/edit documents | No | Yes |
| Make to-do lists | No | Yes |
| Run multiple steps automatically | No | Yes |
| See token usage | No | Yes |
| Permission modes | No | Yes |
The analogy: Browser Claude is like texting someone for advice. Claude Code is like hiring an assistant who sits at your computer and does work for you.
Agent Platform Features You Now Have
Let's go through what's available to you now that you're on an agent platform.
Feature 1: Permission Modes
Remember we enabled "Bypass Permissions" during setup? Here's what that means.
The three modes:
| Mode | What Happens | When to Use | Keyboard Shortcut |
|---|---|---|---|
| Ask Before Edits | Claude asks permission for every file change | Learning, sensitive work | Shift + Tab to cycle |
| Plan Mode | Claude creates a plan first, you approve, then it executes | Complex multi-step tasks | Shift + Tab to cycle |
| Bypass Permissions | Claude does everything automatically | Trusted, repetitive tasks | Shift + Tab to cycle |
To change modes: Look at the bottom of the Claude Code panel. Click the mode indicator or press Shift + Tab to cycle through.
Why Bypass is useful: It gets annoying to approve every little action. "Can I search the internet?" "Can I read this file?" "Can I write this line?" Bypass mode lets Claude work autonomously.
The risk: Claude could theoretically delete files or make changes you didn't want. Use Bypass when you trust the task.
Feature 2: File System Access
This is the big one. Claude Code can:
- Read any file in your workspace
- Create new files
- Edit existing files
- Delete files (if you let it)
The "Copy Path" workflow:
- Right-click any file → Copy Path
- Paste the path in your prompt
- Claude can now read and work with that file
This is how you give Claude context. This is why it could write to your test.md file.
Feature 3: To-Do Lists
When you tell Claude to "make a comprehensive to-do list for yourself," it actually does it.
The consumer app can't do this. It has no concept of task management.
Agent platforms can:
- Create task lists
- Track progress
- Mark items complete
- Self-organize complex work
This is why you should always ask for a to-do list on complex tasks. It makes Claude work more systematically.
Feature 4: Token Window & Compact
Here's something powerful: on an agent platform, you can actually see how much of Claude's memory you're using—and you can manage it.
The value first: When you're working and approaching your limit, you'll see a warning and a "Compact" button appear. Click it, and Claude summarizes the conversation, clears the detailed history, and gives you fresh space to keep working.
The browser app? You just hit a wall and start over. No visibility, no control.
Now let's zoom out—what's actually happening here?
What are tokens?
Tokens are how AI measures text. Roughly 1 token = 3-4 characters. When you type "Hello, how are you?" that's about 6 tokens.
Why do LLMs have token limits?
Every AI model has a "context window"—the amount of text it can consider at once. Think of it like working memory. The model needs to hold your entire conversation in mind to give you a coherent response. That takes compute power, and there's a limit to how much any model can hold.
Different models, different windows:
- Claude (Opus 4.5): 200,000 tokens (~500 pages)
- Gemini 3 Pro: 1,000,000 tokens (~2,500 pages)
- GPT 5.2: ~256,000 tokens (~640 pages)
Bigger token windows—pros and cons:
| Pros | Cons |
|---|---|
| Process entire books or codebases at once | More expensive to run |
| No need to break up massive documents | Can be slower |
| Great for: legal contract review, full codebase analysis, long research papers | May lose focus on details buried in the middle of huge contexts |
Smaller windows—pros and cons:
| Pros | Cons |
|---|---|
| Faster, cheaper | Need strategies for large projects |
| Often more focused on the task at hand | Can't fit everything in one session |
| Great for: focused tasks, iterative work, most day-to-day knowledge work | Requires compacting or handoffs (we'll cover this) |
The takeaway: Different sizes for different jobs. You'll develop a feel for when you need more context vs. when a focused session works better.
We'll go deeper on managing your token window in Section 2.2—including strategies for when you hit the limit and how to avoid the "telephone game" (where context gets distorted as it's compressed and passed along).
What counts toward your token window?
- Everything you type
- Every file you paste in (when you Copy Path a document, that whole document counts)
- Everything Claude responds with
- The entire conversation history
Think of it as a conversation memory. Claude can hold about 500 pages worth of back-and-forth. After that, it starts "forgetting" earlier context.
Why this matters for how you work:
- Big files eat tokens fast (a 50-page transcript = ~15,000 tokens)
- Long conversations accumulate
- You need strategies for large projects (we'll cover these in Section 2.2)
The Compact button is a tool in the tool belt for "token management" we'll talk more about that
Feature 5: Sub-Agents
Here's something you won't find in the browser app: Claude can spawn copies of itself.
When a task is too big or has multiple independent parts, Claude can create "sub-agents"—separate instances that each get their own fresh 200,000 token window. They work in parallel and report back.
Example: You need to research 10 competitors. Instead of doing them one-by-one (slow, uses up your tokens), Claude spawns 10 sub-agents that each research one competitor simultaneously.
We'll go deep on this in Section 3.2. For now, just know it's possible—and it opens up a lot for larger projects.
Quick Recap
You're now operating on a different kind of platform than most people using AI.
| What You Have | What It Means |
|---|---|
| Permission modes | You control the autonomy level |
| File access | AI can read/write your actual work |
| Token window + Compact | You can see and manage your context |
| To-do lists | AI can self-organize complex tasks |
| Sub-agents | AI can parallelize large projects |
Next: Now that you understand the platform, let's get better at communicating with it.