From Prompts to Harnesses: How AI Engineering Has Grown Up

AI engineering has gone through three stages, and knowing where you are in that progression tells you where to put your energy next.Prompt engineering is about how you ask the question. Better wording, examples, and step-by-step instructions get better single answers. It works well for one-off tasks but falls apart when you need consistency across sessions or teams.Context engineering, a term Andrej Karpathy put a name to in December 2025, is about what information the model gets. System prompts, conversation history, retrieved documents, and tool access all shape what the model can see and reason with. Model Context Protocol (MCP) is the infrastructure layer that standardizes how AI tools connect to that information.Harness engineering is the newest piece and the one most teams are missing. Coined by Mitchell Hashimoto in February 2026 and expanded on by OpenAI days later, the idea is this: every time an AI makes a mistake, you change the environment so that specific mistake cannot happen the same way again. The harness is everything around the model: constraint files like AGENTS.md, linters that enforce architecture rules, test gates that block bad output, recovery logic that handles failures, and a separate review step that checks the work with fresh instructions. OpenAI built a million-line codebase this way, with zero manually written code, using a three-person team over five months. Anthropic published their own pattern for long-running agents using an initializer agent that sets up a progress file and feature list, and a coding agent that works one feature at a time.The formula: Agent = Model + Harness. The model is increasingly a standard part. The harness is where the reliability lives.