HuggingFace Just Wrote the Vocabulary We've Been Using: Agent = Model + Harness

On May 25, HuggingFace published an Agent Glossary. The headline definition is one line. “Agent = Model + Harness.” Sergio Paniego and Aritra Roy Gosthipaty wrote it because, in their own framing, the confusion at ICLR 2026 about overlapping terms had become operational debt for the field. They wanted a shared vocabulary. They wrote one.

For anyone selling, buying, or governing AI agents, this is the most important publication of the month. Not because the definitions are novel. We have been using these exact terms for over a year. But because a major model lab now owns the canonical reference. When a procurement team Googles “what is an agent harness,” HuggingFace is what they land on. The vocabulary is now neutral ground.

That changes the conversation. Specifically, it changes who has to do the translation work.

The Definition That Settles a Year of Argument

HuggingFace’s core decomposition is precise enough to put in a contract. The scaffold is the behavior-defining layer: the system prompt, the tool descriptions, the response parsing logic. The harness is the execution layer: the code that calls the model, handles tool invocations, and decides when to stop. The model sits inside. Everything else is scaffolding or harness around it.

The line that does the heavy lifting is the one Paniego and Gosthipaty land on directly: “If you are not the model, you are the harness.” That sentence is procurement-grade. It collapses every fuzzy product category into a binary. Either you ship the weights or you ship the code that orchestrates them. There is no third category.

Sub-agents get the same treatment. A sub-agent has reasoning capability, which is what distinguishes it from a tool (a function call) or a skill (packaged knowledge). Tools execute. Skills inform. Sub-agents decide. Once you accept that taxonomy, the layer at which a vendor competes becomes unambiguous. You can map any AI product to one of those four roles in under thirty seconds.

That mapping is the new procurement skill.

Why the Lab Endorsement Matters More Than the Terms

We have written about this decomposition under the harness definition and traced its cross-discipline applications. The terms are not new. Anthropic uses them in their applied research posts. Practitioners on Twitter use them. The internal documentation of every serious agent team uses them, with local variations.

What was missing was a single citation that a buyer could send to a vendor without it reading as advocacy. If we sent a prospect the Victorino post defining harness, the implicit subtext was “adopt our framework.” If we sent them the Anthropic post, the subtext was “adopt the framework of the lab whose model you might also be considering.” HuggingFace’s position in the ecosystem is closer to neutral. They host models from every lab. They are infrastructure, not a competitor. Their glossary reads as the field defining itself.

That neutrality is what makes the glossary procurement-grade. A CIO can now require, in an RFP, that vendors describe their offering as model, harness, scaffold, sub-agent, tool, or skill, and cite the HuggingFace definition as the reference. The vendor cannot argue with the source. The vendor has to translate their marketing into the glossary’s terms.

This is the moment the vocabulary becomes a procurement weapon rather than an internal tool.

The Three Buyer Questions That Now Have Clean Answers

Before May 25, three questions kept coming up in vendor evaluations and produced muddled answers every time. Each one now has a clean form.

The first is “what layer are you selling at?” Before the glossary, vendors would say “we are a full agent platform” or “we are an agent framework.” Both phrases were marketing, not architecture. With the glossary, the question becomes: do you ship the model, the harness, the scaffold, or some combination? A vendor that cannot answer that question in one sentence is selling you a category, not a product.

The second is “what happens at the seams?” If a vendor sells the harness, what model assumptions does the harness make? If they sell scaffold (a set of prompts and tool descriptions), what harness assumptions does it require? The glossary makes the seams visible. The contracts can now specify which side owns each one.

The third is “where does your governance live?” Most agent governance is implemented in the harness, because that is the layer that decides when to call a tool, when to stop, and how to log. The scaffold can encode intent, but the harness enforces it. Once a procurement team understands that, the security review changes shape. Instead of asking the vendor “is your platform safe,” the team asks “show me the harness behaviors that enforce policy and the scaffold patterns that declare it.” Two specific deliverables instead of one vague one.

Each of these three questions used to require a 30-minute explanation of vocabulary before the substantive answer could begin. The glossary removes the preamble. The conversations get shorter and sharper at the same time.

What Stops Being a Coherent Unit

The hidden move in the HuggingFace post is that “agent” stops being a unit you can buy or evaluate. An agent is a composition. The model is bought from a lab. The harness is bought from a platform vendor or built in-house. The scaffold is written by the team using the agent. The tools are integrated. The skills are curated.

When someone says “we are evaluating Vendor X’s agent,” they are now committing a category error. Vendor X sells one or two layers. The agent only exists when all five layers are assembled. The evaluation has to happen at the layer level.

This will be uncomfortable for vendors who built their pitch around “the agent.” It is liberating for buyers who needed a framework to decompose what they were buying. The leverage shifts toward the buyer who can map a vendor pitch onto the glossary in real time.

What to Do This Week

Take the HuggingFace glossary URL and put it in your next three vendor RFPs. Specifically, in the architecture section, add: “Per the HuggingFace Agent Glossary, identify which of the following layers your offering provides: model, harness, scaffold, sub-agent, tool, skill. For each layer you provide, describe the interface assumptions made about the adjacent layers.”

Then, before the next vendor call, spend ten minutes mapping the vendor’s marketing site to those layers. Note which layers are explicit, which are implied, and which are unclear. The unclear ones are your opening questions. You will find that most pitches collapse two or three layers into one fuzzy term, and that the right discovery question is simply “which layer are you talking about right now?”

If you ship agents internally, mirror the same exercise on your own architecture. Write the one-pager that maps each layer of your stack to the glossary terms. Circulate it to the security, platform, and product teams. The first version will surface three places where two teams meant different things by the same word. Fixing those is the compounding value of adopting the vocabulary.

The vocabulary war is over. HuggingFace called the terms. The teams that adopt them first will run cleaner procurement, cleaner security reviews, and cleaner internal handoffs for the next eighteen months. The teams that keep using “agent” as a unit will spend that same eighteen months explaining what they mean.

This analysis synthesizes Agent Glossary: harness, scaffold, and the AI agent terms worth getting right by Sergio Paniego and Aritra Roy Gosthipaty (HuggingFace, May 2026).

Victorino Group helps procurement and engineering leaders translate the harness vocabulary into RFP-grade evaluation criteria for AI agent vendors. Let’s talk.