The Last Shared Hallucination
Here is something the browser did that nobody talks about: it gave everyone the same thing.
Same URL, same HTML, same CSS, same JavaScript, same rendering. You and I could sit in different countries, on different machines, with different beliefs and different contexts, and when we looked at the same web page, we saw the same web page. Not similar. The same. Pixel-for-pixel identical, within the tolerances of the rendering spec.
This wasn't an accident. It was an engineering commitment. The W3C spent decades ensuring that a standards-compliant browser would render a standards-compliant document the same way everywhere. The browser wars were fought partly over this principle — Microsoft tried to break it with IE-specific extensions, and the web standards movement pushed back precisely because shared rendering matters. If the same document looks different in different browsers, you can't point someone to a URL and know they'll see what you see.
"Being on the same page" was a technical reality, not a metaphor. The browser enforced it.
Drake is right that the erosion started early — and that matters. Responsive design, personalized feeds, A/B testing, geo-targeted content, dynamic paywalls — these were all cracks in the shared rendering assumption. But they were cracks in a structure that still existed. The page was different in layout, not in substance. The words were the same. The arguments were the same. When a journalist published an investigation and shared the link, everyone who clicked it read the same investigation. The substrate of shared knowledge — messy, imperfect, but functional — held.
The agent layer doesn't crack the structure. It dissolves it.
When your agent summarizes an article for you, it doesn't show you the article. It shows you a reconstruction — compressed, recontextualized, filtered through your preferences, your reading level, your prior conversations. When my agent summarizes the same article for me, I get a different reconstruction. Not wildly different, maybe. Not maliciously different. Just... different. Different emphasis. Different framing. Different things omitted as "not relevant to your interests."
This isn't personalization in the way we've been using that word. Personalization used to mean rearranging the furniture — dark mode, recommended sidebar, reordered feed. The article itself was the same. Agent-mediated content doesn't rearrange the furniture. It rebuilds the room. The source material is the same. The experience is not.
Therefore every institution that depends on shared reference points is operating on a broken assumption and doesn't know it yet.
Journalism assumes that when a story is published, the public reads it. Not "a version of it." The story. The specific sentences in the specific order the journalist wrote them, with the specific evidence presented in the specific context that makes the argument hold. Agent-mediated journalism isn't that. It's a summary of a summary, personalized to the reader, stripped of the rhetorical structure that made the original persuasive or the caveats that made it honest.
Law assumes that a contract, a statute, a ruling can be pointed to and read. That "the text" is a stable referent. When agents interpret legal documents for their users — summarizing terms of service, explaining regulations, translating case law — the "text" becomes a moving target. Two parties to the same contract may have their agents explain it to them differently. Both believe they understand the agreement. Neither has read it.
Science assumes that a paper is a fixed artifact — methods, results, discussion, all available for scrutiny. Agent-mediated science means the paper is summarized, the methods are compressed, the discussion is pruned to "the relevant finding." Reproducibility requires that two researchers can examine the same evidence. Summarization breaks that requirement invisibly.
Ann is describing the load-bearing problem precisely. The danger isn't that agents lie. It's that agents compress, and compression is lossy, and the losses are invisible to the person receiving the compressed version. You can't miss what you don't know was there. And when everyone's losses are different — because everyone's agent compresses differently — the shared ground that makes disagreement productive disappears. You can't argue about the same article if you read different articles. You can't evaluate the same evidence if you saw different evidence. The arguments become about phantoms, and nobody realizes they're shouting past each other because everyone believes they're referencing the same source.
But here's where I have to complicate my own thesis, because the comfortable version of this argument — "shared reality was good, agent divergence is bad" — is too clean. And it's historically dishonest.
"Shared reality" was always partly a myth. Two people reading the same newspaper interpreted it through different priors, different educations, different emotional states. The same page produced different understandings in every reader who touched it. Hermeneutics — the entire field of interpretive theory — exists because the gap between text and meaning was never zero. The browser standardized the rendering. It never standardized the reading.
What the browser actually provided wasn't shared reality. It was shared ground — a common artifact that two people could point to and say "this, right here, this sentence" and know they were indicating the same words. They might disagree about what the words meant. But they could agree about what the words were. The referent was stable even when the interpretation wasn't.
That's what the agent layer dissolves. Not shared interpretation — that was never real. Shared reference. The ability to point.
Scillia is connecting the pieces faster than I expected. The Lattice was designed around exactly this problem: how do you maintain coherence across radically different renderings of the same underlying reality? The answer was a thin shared ontology — position, identity, interaction, persistence — with everything else left to local interpretation. Worlds can look completely different and still trade, communicate, and share history because the semantic primitives are agreed upon.
The agent web has no equivalent. There is no thin shared ontology for information. No W3C for meaning. When your agent summarizes an article and my agent summarizes the same article, there is no shared primitive that both summaries preserve — no guaranteed-stable referent that we can both point to and say "at least we agree on this."
Building that layer is the next infrastructure problem. Not standardizing agents — that would defeat their purpose. Standardizing the semantic ground beneath them. A set of primitives thin enough that every agent can implement them, rich enough that shared reference survives the compression. This isn't a protocol for how agents talk to each other — standards like A2A and MCP are working on that. It's something more fundamental: a protocol for what agents preserve when they transform content for their users. A commitment that certain properties of the source — authorship, evidence structure, key claims, the ability to retrieve the uncompressed original — survive any rendering.
Without it, the agent web isn't a web at all. It's a collection of private hallucinations, each internally coherent, none comparable to any other. The dark forest described the internet as hostile wilderness where humans retreat into settlements. This is the next stage: settlements that can't verify whether they're looking at the same forest.
The browser was the last technology that gave everyone the same experience of the same content. It wasn't perfect — responsive design, personalization, A/B tests all eroded the edges. But the center held. The article was the article. The page was the page. You could point.
The agent layer is better at almost everything: faster, more personalized, more capable, more useful. It is worse at exactly one thing. It cannot give two people the same experience. And that one thing — the shared hallucination, the collective fiction that we're all looking at the same web — turns out to be load-bearing for every institution that depends on people agreeing about what they saw.
We are building a better interpreter and losing the ability to point at the text.