AI Agent = LLM + Harness
- LLM: Claude, ChatGPT, Gemini, etc.
- Harness: Codex, Claude Code, OpenClaw, etc.
Prompt -> Context -> Harness
AGENTS.md: Rules for AI Agent
- Should not very long, more like a map not a dictionary Nature Language Harness
Life-long AI Agent: Feedback
- Ground truth: As close to ground truth as possible
- Numerical Feedback: As large reward possible
- Verbalized Feedback: "good job", "you are stupid", etc.
- Environment Feedback: Error message from program