← Back to digest
AI Agents

"Harness Engineering" hits HN

What's this about?

A dense writeup framing harness design as an engineering discipline reached the HN front page. The thread that followed is less about the content than about whether the vocabulary means anything.

What is a harness, in one sentence?

The code around the model — the loop, the tools, the memory, the planning — that turns a text predictor into something that completes tasks. See the JIT-Agent entry above for why it may matter more than the model.

What's the community saying?

HN 118 points, 42 comments — a genuinely split thread
  1. The most substantive observation, from someone who liked it: agent review is listed as an explicit deprecation target, meaning a harness matures by getting *dumber and cheaper*, not smarter.
  2. The strongest objection is about provenance — the author runs a consultancy, the README is very long, and several readers conclude it was model-written and not really meant for human consumption.
  3. A defence from someone attending the author's workshop: there are useful ideas here and much of the material is free and open-source.
  4. The comparison that stings: "typical material coming out of consulting companies, designed to obfuscate rather than educate."
  5. A recommended alternative gets upvoted — a plainer explanation of what a harness is, posted the previous week.
  6. The philosophical undercurrent: several argue harness, skills, agents and RAG are largely invented words with little semantic content, being retrofitted with meaning after the fact.
← Agentic shopping is unpredictableDeepMind pilots double-blind AI evaluations →