File size: 7,768 Bytes
7d1f7d5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 | # 01 — The Vision: What We're Building & Why
## 🎯 Your Question
> "Manus is amazing. How do they do it? Can we build something like that?"
**Short answer:** Yes! Not identical — Manus has hundreds of engineers and millions in funding. But we can build a **"child version"** that captures the core idea and teaches you every concept along the way.
---
## 🤖 What Is Manus AI?
Manus (acquired by Meta) is an **AI agent** — not just a chatbot. Here's what makes it special:
### 1. It Actually DOES Things (Not Just Talks)
| ChatGPT/Claude | Manus |
|---------------|-------|
| "Here's how to find Python files..." | *Actually runs the command and shows you* |
| "Here's a script idea..." | *Writes, tests, and deploys the code* |
| "I can help you plan..." | *Plans, executes, and verifies* |
### 2. Three Specialized Agents Working Together
Manus uses **three sub-agents** that coordinate:
```
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ PLANNER │────▶│ EXECUTOR │────▶│ VERIFIER │
│ │ │ │ │ │
│ "Break this │ │ "Run shell │ │ "Check if │
│ into steps"│ │ commands" │ │ it worked" │
│ │ │ │ │ │
│ Strategize │ │ Navigate │ │ Quality │
│ multi-step │ │ web, write │ │ control │
│ path │ │ code, use │ │ & fix │
│ │ │ tools │ │ errors │
└─────────────┘ └─────────────┘ └─────────────┘
```
### 3. Persistent Cloud Environment
Manus runs in a **cloud VM** (virtual machine):
- Files persist between sessions
- Can install software (`pip install`, `npm install`)
- Works while you sleep (asynchronous)
### 4. Can Browse 50+ Websites Simultaneously
For research tasks, Manus spawns many parallel agents to gather info.
---
## 🔬 What We're Building: "Mini-Manus"
### Our Simpler Architecture
Instead of three separate agents + cloud VM, we use **ONE model** with a loop:
```
User: "Find all Python files and count them"
│
▼
┌─────────────────────────────────────────┐
│ MCP-Agent-1.7B (Our Model) │
│ │
│ ┌─── THINK ───┐ │
│ │ "I need to │ │
│ │ list .py │ │
│ │ files" │ │
│ └──────┬───────┘ │
│ │ │
│ ┌─── ACT ─────┐ │
│ │ shell_exec({│ ◀── ONE MODEL plays │
│ │ "command": │ ALL three roles │
│ │ "find . │ (planner + │
│ │ -name │ executor + │
│ │ '*.py'" │ verifier) │
│ │ }) │ │
│ └──────┬───────┘ │
│ │ │
│ ▼ (Result: "main.py, test.py, utils.py")
│ │
│ ┌─── VERIFY ──┐ │
│ │ "Got 3 │ │
│ │ files. Now │ │
│ │ count." │ │
│ └──────┬───────┘ │
│ │ │
│ ┌─── ACT ─────┐ │
│ │ python_exec({│ │
│ │ "code": │ │
│ │ "print(3)"│ │
│ │ }) │ │
│ └──────┬───────┘ │
│ │ │
│ ▼ (Result: "3") │
│ │
│ ┌── RESPOND ──┐ │
│ │ "Found 3 │ │
│ │ Python │ │
│ │ files! ✅" │ │
│ └─────────────┘ │
└─────────────────────────────────────────┘
```
### Key Differences from Manus
| Feature | Manus | Mini-Manus (Ours) |
|---------|-------|-------------------|
| Agents | 3 specialized (Planner/Executor/Verifier) | 1 model, all roles |
| Environment | Cloud VM | Local/Gradio Space |
| Parallelism | 50+ simultaneous | Sequential (one at a time) |
| Cost | $$$/month | $3 one-time |
| Model Size | GPT-4 class (100B+) | 1.7B (100× smaller!) |
| Persistence | Files persist forever | Session-based |
| Web Browsing | Real browser | DuckDuckGo search API |
### Why This Still Impresses People
1. **It runs LOCALLY** — No API keys, no cloud costs, no rate limits
2. **It actually DOES things** — Not just text, but real shell commands, file operations, Python execution
3. **It's 100× smaller** than Manus's models but still functional
4. **It's OPEN SOURCE** — Anyone can use, modify, improve it
5. **YOU trained it** — From base model to agent in one project
---
## 🧠 The Core Insight: Why Small Models CAN Work for Agents
You might think: *"How can a 1.7B model compete with GPT-4?"*
The secret is **FOCUS**.
GPT-4 is a generalist — it knows about history, science, poetry, coding, everything.
Our model is a **specialist** — it ONLY knows about tool-calling.
Think of it like this:
- GPT-4 = A professor who can teach any subject
- Our model = A skilled technician who only knows how to use tools
The **TinyAgent paper** proved this: a 1.1B model fine-tuned on tool-calling
data matched GPT-4-Turbo at function-calling tasks. Not because it's smarter,
but because it's **focused**.
---
## 📋 What Makes This a "WOW" Project
When you show this to people, they'll be impressed because:
### 1. "You trained your own AI agent?"
Most people think you need a PhD and a supercomputer. You don't.
### 2. "It runs on a laptop?"
1.7B parameters = 4GB in memory. Runs on any gaming laptop.
### 3. "It can actually modify files?"
Not just text generation — real file system operations, shell commands, Python execution.
### 4. "It costs $3?"
Compared to Manus's pricing (or OpenAI API costs), this is almost free.
### 5. "You built this yourself?"
From research → data → training → app. Full pipeline.
---
## 🎓 What You'll Learn From This Project
By the end, you'll understand:
- ✅ How AI agents work (ReAct pattern)
- ✅ What MCP is and why it matters
- ✅ How to pick base models for different budgets
- ✅ LoRA: the magic of cheap fine-tuning
- ✅ SFT: supervised fine-tuning step-by-step
- ✅ How to tune hyperparameters (learning rate, batch size, epochs)
- ✅ How to build an agent harness
- ✅ How to deploy ML models
- ✅ How to read research papers and apply them
**If you can train a 1.7B model, you can train a 70B model.**
The concepts are identical — only the scale changes.
---
## 🔜 Next Step
Read `02-research.md` to see what papers and datasets we found, and why we made the choices we did.
|