agent-hack

The simplest possible agent: LLM + Tools + Loop

The Bootstrap Challenge: Building Claude Code from Scratch

The Concept

Start with an agent that can only READ files. Use it to teach yourself how to add WRITE. Then use READ+WRITE to add SHELL. This mimics how real development tools evolve!

Setup

1. Install uv:

   curl -LsSf https://astral.sh/uv/install.sh | sh

2. Create .env file:

   echo "ANTHROPIC_API_KEY=your-api-key-here" > .env

3. Paste in your actual Anthropic API key
4. Run:

   uv run agent.py

Stage 1: READ Only (What you start with)

uv run agent.py

Your first conversation:

> Can you read agent.py and tell me how to add a write tool?
[Agent reads the file, sees the pattern, and gives you exact code to add]

Stage 2: Adding WRITE (Manual step)

1. The agent gives you code for a write function
2. You manually paste this into agent.py
3. Update the tools list: tools=[read, write]
4. Restart the agent

Stage 3: Adding SHELL (Automated!)

Now with both read and write:

> Can you read your own file (agent.py) and add a shell tool to yourself?
[Agent reads itself, generates shell function, writes it back to the file]

Stage 4: Victory!

You now have a fully functional agent that you bootstrapped from almost nothing!

Learning Objectives

• Understand how agents work by building one piece by piece
• Experience the limitations before solving them
• See how tools enable other tools
• Learn the agent's code patterns by using the agent itself

Bonus Challenges

Once you have all three tools:

• Build something that is actually cool using it... maybe with the GenAI tooling?
• Use it as a CLI tool

The Meta Lesson

You're using an AI agent to help you build an AI agent. This is how development works - we use tools to build better tools!