Word of the Year: Harness
If you spent any time around AI tooling this year, you watched “harness” go from a word nobody used to one in every changelog, blog post, and conference slide. Anthropic’s own docs describe Claude Code as “the agentic harness around Claude.” A year ago that sentence would have needed a footnote. Now it ships in the product copy.
Most AI words that catch on this fast arrive pre-hollowed, drained of meaning before you even notice them. Harness still means something, at least for now. You can write the definition on an index card, hand it to a coworker, and they will be more useful five minutes later. That is rare enough in this field to be worth celebrating before the marketing departments find it.
What it actually means
The cleanest version comes from Viv Trivedy and Will Brown: Agent = Model + Harness. If you are not the model, you are the harness. The model takes text in and produces text out. On its own it answers one prompt and stops, with no memory of the last call and no way to run a tool even when it badly wants to. The harness is everything wrapped around the model that turns it into something useful: the prompt and tools, the context management, the loop that keeps it going, and the guardrails that stop it doing anything too stupid.
Simon Willison has the shortest definition of the active part: an LLM agent runs tools in a loop to achieve a goal. Claude Code, Codex, Cursor, and the 50-line Python script you wrote last Tuesday are all harnesses. Some are elaborate and some fit on a napkin, but they are the same category of thing. Swap a better model into the same harness and the behavior shifts. Rebuilding the harness around the same model shifts it more.
You can use this to reason about why two products running the same underlying model feel completely different, or why your own setup got noticeably better over a month when the model never changed at all.
The graveyard
Now the bad news, which is that I have seen this movie before. AI has a habit of taking a precise word, feeding it to the marketing machine, and handing it back as noise.
“Agentic” is the freshest corpse. It started as a reasonable adjective in a handful of research papers, and inside about two years it got bolted onto calendar apps and, somewhere out there, an air conditioner. Gartner has a name for it: agent washing. By mid-2025 it was looking at the thousands of vendors selling “agentic AI” and finding that about 130 of them were the genuine article. It also expects more than 40 percent of agentic AI projects to be canceled by the end of 2027. The line that turns up in every honest write-up by now is that when everything is agentic, nothing is.
“AI” itself is further gone, so far gone that Apple agreed to a 250 million dollar settlement over false advertising of Apple Intelligence features. For most buyers “AI-powered” now means “we added an algorithm,” and the average person can no longer tell a real model from a rules engine in a chatbot costume. The research crowd did the same thing to “AGI,” which in 2026 has serious people disagreeing by a decade about whether we already have it, since no two of them share a definition. “Vibe coding” did not even survive a year before it stopped meaning the one specific thing it was coined for.
None of these words got killed on purpose. A word that does real work gets valuable, and anything valuable pulls in people who want the credibility without doing the work. They wear it out, and after a while it stops telling you much of anything.
The quiet one
“Reasoning” is the dilution that annoys me most, because it is anthropomorphic and almost nobody flinches at it. In the wild “reasoning” now describes any model output longer than a sentence. “Chain-of-thought” sounds like a mind working through a hard problem. What it names is a model generating more text before it gives the answer.
A philosophy paper at FAccT 2026 has a sharp word for this trick: glosslighting. You reach for a human-loaded term like reasoning or hallucination or introspection, lean on everything the word implies about minds, and then retreat to the narrow technical meaning the moment somebody pushes back.
This one matters more than the rest for anyone actually building. Once “reasoning” means “produced some text,” you have lost the words you needed to explain to a stakeholder that the model is not, in fact, reasoning about their particular problem. The words you would reach for to name the limitation are the same ones the marketing team already used up, so you end up explaining it from scratch every time someone asks.
Why it keeps happening
None of this needs a conspiracy. There are tens of thousands of AI companies and maybe a dozen words that sound impressive in a funding round, so the words get reused until they wear through. Every press release has to sound like a breakthrough, and the supply of real breakthroughs does not keep pace with the demand for announcements.
There is a real cost under the comedy. Psychologists call it semantic satiation when you repeat a word so many times it briefly stops meaning anything, and then your brain resets. The industry’s version does not reset. The words come out the far side permanently hollow, and you lose the ability to describe the real thing on the day it finally shows up.
Harness is next
So treat this partly as a warning label, because harness is standing roughly where “agentic” stood two years ago. The cracks already show if you look. People cannot agree whether a harness is the same thing as a “scaffold,” or how either of those relates to a “framework” or a “runtime.” The HuggingFace glossary spends a whole section separating harness from scaffold, and the harness-engineering guides pile on framework and runtime, which is usually the first sign a word is about to slip its moorings.
The distinctions are worth holding onto anyway. The harness is the orchestration that wraps a model into an agent: the loop that keeps calling the model, plus the context handling and tool execution around it. Scaffold usually points at the static setup the model works from, its system prompt and tool descriptions, though plenty of people use it interchangeably with harness and will argue the point. Frameworks like LangChain or CrewAI hand you parts to build one with. The runtime is whatever keeps the whole thing alive and running. They are not interchangeable, even if people treat them that way. That holds until some vendor ships “agentic harness-powered” smart luggage.
So enjoy “harness” while it still means something. The species is not extinct yet, but it is on the list, and you know how these things go.