Ryan Lopopolo of the OpenAI Codex team: "Agents are not hard; the harness is hard." Here's what actually makes up that harness, and why the prompt is one small piece of it: → tools: capabilities to actually act on real systems, not just describe actions → memory and state: persists across steps and sessions instead of resetting every run → constraints: deterministic guardrails that block dangerous actions instead of politely asking the model not to take them → verification: a separate check catching mistakes, since models cannot reliably evaluate their own work → orchestration: the logic sequencing everything and deciding when to stop The field moved through three stages to get here. prompt engineering in 2023, context engineering in 2024-2025, harness engineering now. each layer nested inside the last instead of replacing it. A prompt is one line of the program. The harness is the whole program. Bookmark so you do not lose it! Follow @neil_xbt for more AI content!