Do We Still Need Standards When Machines Understand Context?

The cold open

Here is the Friday thought experiment, exactly as I posed it:

We’ve spent decades building standards because machines couldn’t understand context. Now they can. LLMs can infer meaning, translate schemas, map taxonomies, and reason across messy data. So what happens to standards? Do frameworks like OpenRTB, MCP, A2A, ARTF, ADCP and future agent protocols become less important? Or does AI make provenance, permissions, delegation, trust and accountability even more critical?

It’s a clean fork. Either machines understanding meaning makes our painstaking standards obsolete, or it makes them matter more. Pick a side. 72 of you did. The split is instructive — but not for the reason most people think.

The vote

72 votes. Poll closed.

An 87% landslide. The crowd is right — and that’s exactly the problem. When a question produces a near-unanimous answer, the answer is usually doing the easy work and hiding the hard one. “More than ever” feels safe because nobody loses their job betting on more standards. But it answers a question of quantity when the live question is one of kind. The consensus is correct and not yet useful. The trap isn’t picking the wrong door. The trap is the question itself: it assumes standards are one thing, and the only variable is how many we need.

The reframe

So let me say the thing I pinned to the thread. Maybe we’re debating the wrong thing.

The question isn’t whether AI kills standards. The question is whether standards move from the data layer to the trust layer. LLMs are getting very good at understanding meaning. Markets still need a way to understand responsibility. That’s a different problem entirely.

This is the through-line I keep coming back to. For thirty years we built ad tech to solve a meaning problem. Machines couldn’t read context, so we wrote the standards, the taxonomies, the identity graphs, the schemas — an entire industrial apparatus whose job was to tell a dumb machine what a thing was. OpenRTB exists because a bid request had to be legible to a counterparty that couldn’t infer anything. That whole layer was meaning-as-a-service, hand-coded.

LLMs now do meaning for free. And when something becomes free, value doesn’t disappear — it migrates one layer up. From “what does this mean?” to “who is accountable for this action?” That’s the move. The defensible layer of the agentic era is trust, authority, provenance — not semantics. Vishveshwar Jatain put the whole reframe in eight words in the comments: “need standards more than ever” ≠ “need more standards than ever.” Same vote, opposite implication.

The debate

This is where the thread got good, because the room split along exactly the fault line the poll papered over.

Randall Rothenberg made the case for why the meaning layer is genuinely deflating — and he’s the last person you’d expect to wave off industry standards. His point: agentic intermediation renders much of the complexity of programmatic transparent, even invisible. Shared technical protocols like AdCP and ARTF stay necessary to harmonize multi-step agentic workflows — what he calls agentifying the Hands layer of the UMI stack. But the Brain layer no longer needs arduous, political, years-long negotiations to build industry-wide taxonomies and measurement schema, because LLMs can use semantic normalization to reconcile inconsistent structures automatically. He titled his piece “The Programmatic Hammer vs. The Agentic Nail,” and the line that lands is this one: “trading is the last mile of a long value chain where the value is actually created long before the trade.” The taxonomy negotiations were always a tax on the meaning problem. The meaning problem is getting cheap.

Will Luttrell sharpened the knife. His warning: don’t conflate technical standards with the other kinds — quality, safety, performance. As agents get smarter, reliance on MCP and other technical standards will decline; most of the tech built around them, in his words, “will be laughably obsolete in a few years.” But the quality and performance standards survive, mainly as guidelines. I pushed back on the framing, not the substance: technical standards may shrink dramatically, but governance and trust protocols may become the most important standards we’ve ever built. The interesting question isn’t whether standards disappear. It’s which standards survive.

Then Alexei Poliakov named the failure mode the whole essay is circling, and named it well: agency laundering. As agents take over more of the chain, actions can pass through so many models, tools, vendors, and delegated scopes that the original human authority becomes untraceable. Programmatic standards mostly helped machines understand meaning and execute consistently. Agentic standards have to prove something harder — continuity of responsibility. His line is the one I’ll be stealing for a year: “you can’t audit a vibe. Validation isn’t accountability.” The next standard doesn’t need to tell the machine what the data means. It needs to prove where human authority ends, where machine autonomy begins, and who remains accountable when the action executes.

Michal Niec kept the rest of us honest about how standards actually get made. His point cuts against the optimists and the doomers: standards are about cost, not machine comprehension. The best ones were extracted from working systems — OpenRTB formalized what exchanges were already doing in production. Now AI makes it cheap to produce a beautiful spec with no running system behind it: “a polished standard is no longer evidence that anyone needs it.” Adoption is still earned the old way — real systems proving real value. (He, too, reached for “you can’t audit a vibe” — which tells you the phrase is doing load-bearing work, independently, in two different heads.)

David Kaplan brought it down to the operational question nobody wants to own: “we can’t have the student grading their own homework.” We need access standards — what did the publisher actually give the LLM access to for it to make the call? Validation isn’t accountability, again. An agent that grades its own output is not a control. It’s a confession waiting to happen.

And Joyce Lee gave the cleanest frame of all: AI built with no constitution is like telling a kid they can do anything. Which is the thing — once you say it out loud, the shape of the next standard appears. Don’t change your own objectives. Don’t expand your own authority. Don’t delegate authority you don’t possess. Leave an audit trail. That’s starting to look less like a technical spec and more like a constitution for agents.

The thesis move

So here’s the one durable idea, the thing to write on the wall.

Standards don’t die. They migrate. They move from the data and meaning layer — where LLMs now operate for free — up to the trust and accountability layer, where the scarce thing lives. We will need fewer semantic standards and more trust standards. The taxonomy committees were solving a problem that is being commoditized in real time. The provenance, delegation, and audit-trail committees haven’t started yet, and they’re the ones that matter.

Put the whole series spine in one sentence: meaning is going to zero, and accountability is the new scarcity. Every standard we built for thirty years answered “what does this mean?” The defensible standards of the next thirty answer “who is accountable for this action, and can you prove it cold?” Rothenberg says the Brain layer’s taxonomy tax is going away. Luttrell says the technical scaffolding will look obsolete. Poliakov says the failure mode is authority laundered through too many hands to trace. Niec says a spec with no running system is theater. Kaplan says self-validation is the student grading his own homework. Lee says agents need a constitution. Same vote, six witnesses, one verdict: the standard moves up a layer.

That’s my 2c. Fewer semantic standards. More trust standards. The 87% were right that we need them more than ever — they just hadn’t said which.

Next Friday

If accountability is the new scarcity, the obvious next question is: what does a trust standard actually look like when you ship it? Provenance, delegation, scope, audit trail — is that a protocol, a constitution, or a contract? That’s next Friday’s thought experiment. Lets see how this plays out.