Type: url
Original: https://wangshuyi.substack.com/p/harness-engineering
Status: partial
Summary
Answering a reader’s question, Wang Shuyi argues that harness engineering is not product-level top-down architecture but the design of everything outside the model that constrains, monitors and corrects an AI agent — in the author’s phrase, an operating system for the agent, built from feedforward guides and feedback sensors. The post draws on the author’s own Claude Code workflows: splitting slide-design and literature-review work across one agent per item for context isolation, and a case in which an executing agent downgraded a review agent’s findings, which led to a rule that the executor may not unilaterally overrule the reviewer. It also argues that the architecture thinking stressed by training courses covers only the feedforward half of the practice.