I had a folder on my laptop called agent-setup/ . Inside: a CLAUDE.md, a couple of YAML configs, three skills, and a bash script wiring them together. For six months I called it "the setup." When teammates asked what I was doing this quarter, I'd say things like "rebuilding the setup" or "the setup ate my Tuesday." Nobody pushed back. Engineers tolerate vague vocabulary the way doctors tolerate bad handwriting. Then in March, an arXiv paper landed titled Natural-Language Agent Harnesses . I read the abstract, looked at my agent-setup/ folder, and felt the specific embarrassment of someone who just learned the proper word for a thing they'd been mispronouncing in public for a year. The paper is arXiv 2603.25723 , submitted March 26, 2026 by Linyue Pan and four colleagues from Tsinghua and Harbin Institute of Technology. It's the first time I've seen the harness concept treated as a first-class research object instead of a footnote in someone else's agent framework.…