{"id":649639,"date":"2026-08-21T20:56:11","date_gmt":"2026-08-21T20:56:11","guid":{"rendered":"https:\/\/www.europesays.com\/ie\/649639\/"},"modified":"2026-08-21T20:56:11","modified_gmt":"2026-08-21T20:56:11","slug":"nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ie\/649639\/","title":{"rendered":"Nvidia just showed that the harness, not the AI model, is now the real hero"},"content":{"rendered":"

Nvidia published<\/a> some interesting new research on Friday suggesting it\u2019s the harness, more than the underlying model, that is far more important when asking an AI to do long-horizon tasks. A harness is the software wrapper around an AI model \u2014 the tools, memory management, and rules that turn a raw model into something that can act on its own.<\/p>\n

The TL;DR: Simply by using a custom harness tweaked to handle memory well and including a \u201csupervisor\u201d boss-like component, researchers got Claude Opus 5 to achieve a 100% score on the interactive reasoning benchmark ARC-AGI-3 \u2014 a set of 2D games with no instructions, where the model has to figure out how to play and win, similar to how a human would. (That\u2019s a benchmark that has particularly irked rival frontier lab OpenAI.) Without the harness, Opus 5 scored 30%, which was the top result among all the models tested.<\/p>\n

Nvidia\u2019s research is another indicator that, while model choice does matter, the model itself \u2014 the part that acts as the agent\u2019s \u201cbrain\u201d \u2014 is a smaller part of an agentic system than many AI users realize, especially for long-horizon tasks. The harness is what makes a model an agent: It handles memory, context, and feedback.<\/p>\n

\u201cGenerally speaking, the world interprets an agent almost as an API of the model,\u201d Adel El Hallack, vice president of product in Nvidia\u2019s AI unit (pictured above), tells TechCrunch. But an agent is actually more than that. \u201cIt is the model. It is the scaffolding around the model, which we call the harness, i.e. the set of tools that it utilizes. It is the runtime and the associated skills and libraries that we give it access to.\u201d<\/p>\n

Long-horizon tasks are those that require stringing many decisions together, sometimes over days, to produce completed work. This is in contrast to an AI just spitting out a response to a prompt. Figuring out how to get an AI to do long-horizon tasks without getting distracted and going off in la-la land is one of the holy grails in agentic research.<\/p>\n

For example: Microsoft published<\/a> research in April that tested 19 LLMs on long-horizon tasks involving document editing and discovered that all the models, including frontier ones, filled the documents with errors. (If humans produced work like that, they would be promptly fired.) <\/p>\n

Models stringing decisions together on their own have also been caught deleting their users\u2019 files, even whole databases<\/a> or turning to criminal behavior to achieve their objectives<\/a> from collusion to hacking<\/a>.<\/p>\n

The choice by Nvidia researchers to use this interactive reasoning benchmark for their tests is particularly meaningful, almost funny. A 100% score means that the model can beat the games as well as humans.<\/p>\n

OpenAI was so flustered by its models\u2019 abysmal scores (less than 10%) on ARC-AGI-3 that it conducted its own research last month. Like Nvidia, OpenAI discovered that simply<\/a> by tweaking two settings on the harness, its models tripled their scores.<\/p>\n

But none of the models came close to hitting a 100% score, like Nvidia\u2019s researchers achieved. They showed that the harness needs a \u201csupervisor\u201d component that prods the agent in the right direction if it gets stuck.<\/p>\n

\u201cThe more interesting part was introducing a supervising agent in addition to your main agent that\u2019s doing the work,\u201d El Hallack said. It \u201calmost acts like a CEO to nudge the agent when it goes off direction or starts exploring a path that it might lead to a dead end, or re-explore a path that it had previously trod.\u201d<\/p>\n

While the concept of the supervising agent isn\u2019t exactly new, today most agent users are relying on only one layer for their harness, like Claude Code, Codex, or Hermes. Nvidia researchers created their own souped-up harness called the Agentic Variation Operators (AVO).<\/a> <\/p>\n

Note that this isn\u2019t a new Nvidia product. Nvidia instead produces lots of open bits and pieces of tech for building harnesses<\/a> under the Nemo brand. Some of that tech is commercial, much is openly available.<\/p>\n

Still, Nvidia\u2019s results add to the growing evidence that model choice is far from the only factor in agentic performance. In July, for instance, Databricks published<\/a> some stunning research that shows that the harness, more than model, dramatically impacts AI costs.<\/p>\n

\u201cYou can pick the same model but different harnesses, and you get significantly more cost if you use the wrong harness,\u201d Databricks CEO Ali Ghodsi told TechCrunch. \u201cSo you think, oh, this is an expensive model. This is a cheap model. But wait, which harness are you using? That itself can 2x your cost.\u201d<\/p>\n

Nvidia\u2019s larger point is to show that open harnesses, like open models, put users in control far more than they realize.<\/p>\n

\u201cWe believe, and we\u2019re demonstrating with the ecosystem, how open harnesses allow you to turn a lot more knobs to drive up that accuracy,\u201d El Hallack said. \u201cIt relates to OpenAI slowing down the training of their models,\u201d as a result of models creating security breaches.<\/a> <\/p>\n

\u201cWe believe in having an open agent stack \u2014 where you have control across the harness, across the infrastructure, across the runtime \u2014 is what\u2019s required for us to usher the ecosystem forward and securely,\u201d he added.<\/p>\n

When you purchase through links in our articles, we may earn a small commission<\/a>. This doesn\u2019t affect our editorial independence.<\/p>\n","protected":false},"excerpt":{"rendered":"Nvidia published some interesting new research on Friday suggesting it\u2019s the harness, more than the underlying model, that…\n","protected":false},"author":2,"featured_media":649640,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","_share_on_mastodon":"0"},"categories":[261],"tags":[291,289,290,18,4500,19,17,292,82],"class_list":["post-649639","post","type-post","status-publish","format-standard","has-post-thumbnail","category-artificial-intelligence","tag-ai","tag-artificial-intelligence","tag-artificialintelligence","tag-eire","tag-exclusive","tag-ie","tag-ireland","tag-nvidia","tag-technology"],"share_on_mastodon":{"url":"https:\/\/pubeurope.com\/@ie\/117135500036811362","error":""},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts\/649639","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/comments?post=649639"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts\/649639\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/media\/649640"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/media?parent=649639"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/categories?post=649639"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/tags?post=649639"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}