ampcode-how-to-build-agent

How to Build an Agent (Ampcode)

Thorsten Ball (Amp) writes a step-by-step tutorial building a code-editing agent in <400 lines of Go. Subtitle: “The Emperor Has No Clothes.”

Key Claims

• “It’s an LLM, a loop, and enough tokens” — there’s no secret behind agent behavior
• Agent definition: “an LLM with access to tools, giving it the ability to modify something outside the context window”
• Tools are just structured “wink if you want me to X” — models are trained and fine-tuned for this
• The entire agent is: conversation loop → tool definitions → tool execution → result feedback → continue
• A fully functional code-editing agent needs just 3 tools: read_file, list_files, edit_file
• “The rest, the stuff that makes Amp so addictive? Elbow grease.”

Tutorial Structure

1. Basic chat loop (90 lines) — conversation management
2. Tool definitions — name, description, input schema, function
3. Tool use loop — when model’s stop_reason is tool_use, execute and continue
4. Three tools build a capable agent

Takeaways

• The “emperor has no clothes” framing: agents look magical but the core loop is simple
• The server is stateless — all state management is on the client side
• Models are “very eager” to use tools — they “know” they can get more information
• The complexity is in the tooling, prompt engineering, and edge case handling, not the loop itself
• This is the Amp team’s philosophy: simplicity first

Connections

• How to Build an Agent (existing page) — overlapping topic
• Agent Definition — Simon Willison — agent taxonomy
• Coding Agents — the practical pattern
• Harness Engineering — “elbow grease” = harness engineering