Building an AI Agent from Scratch in Rust

Built a minimal AI agent in ~150 lines of Rust to understand the core agent loop pattern. The repo: qepo17/kodi.

The agent loop is surprisingly simple:

1. Send messages to an LLM (via OpenRouter → Claude Sonnet)
2. Check if the response contains tool calls
3. If yes → execute the tools, append results to message history, loop back to step 1
4. If no → print the final response and exit

The architecture breaks down into four clean files:

• models.rs — OpenAI-compatible request/response types (serde-powered)
• clients.rs — The agent loop: send → check tool calls → execute → repeat
• tools.rs — Tool execution (currently just bash)
• main.rs — Entry point, reads API key and query from args

Key implementation details:

• Messages use a tagged enum (#[serde(tag = "role")]) for clean serialization of system/user/assistant/tool messages
• Tool call results get appended as tool role messages with the matching tool_call_id
• The finish_reason field distinguishes between "tool_calls" (keep looping) and "stop" (done)
• Tool arguments come as a JSON string, not a parsed object — you parse them yourself

The takeaway: the “magic” of AI agents is just a while loop with tool dispatch. The model decides what to do, the harness decides how to do it. Once you see the pattern, the whole concept clicks — and you realize you can build one yourself without any framework.