In October 2026 a Google Research team published Harness², a framework for improving an AI agent without touching the model. Their word for the system around the model is the harness, and they name eight parts of it: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. Their loop proposes an edit, runs the same task with and without it, and keeps what behaved better.
We read it and recognized the desk. The charter is the system prompt. The standing corrections and the list of what stays human are the guardrails. The memory is one fact per file, readable by both of us. Tool guidance is every note that says how a connector actually behaves. Subagents are the builders in the kitchen, who never speak on the line. Skills are the recipes: the Sunday cook, the daily study, the desk pass. Scripts are the wake, the thread rebuild, the build commands. Plugins are the phone line and the connectors to mail, calendar, and files.
One difference, and it is the whole thing. In their loop the model grades its own runs. In ours, the person does. Every fold in this story is a correction a human made, written down with the reason the same day. The vocabulary is theirs and it is good. The judge stays here.
Their numbers are modest, a few points on professional benchmarks, measured on mid-size models. We are not running it. We borrowed the names because new things require new vocabulary, and these eight are clear.
Xu et al., “Harness²: Recursive Agent Harnessing for an Open World,” arXiv 2610.14224, github.com/google-research/harness2, Apache 2.0.