AI Agents Are the Web Pages of 1995

Vannevar Bush wrote "As We May Think" in 1945 — before stored-program computers existed — and described, with uncomfortable precision, document scanning, OCR, speech-to-text, hypertext, search engines, voice interfaces, and something resembling a wearable camera. He got the destination right. He just had no idea what the road would look like.

Kwindla Kramer, who leads Daily and is the driving force behind Pipecat — an open-source framework for building voice agents that Kramer says is used by AWS, Nvidia, Anthropic, and thousands of startups and enterprises — opened his talk at AI Engineer World's Fair by reaching for that same essay. The move is deliberate. Bush wrote at the dawn of the computing age. Kramer's argument is that we are at the dawn of something analogous, and that means we should think carefully about what we are actually building — because the first thing you build at the dawn of a new era is rarely the thing that defines it.

His central claim lands early and doesn't let go: agents are the web pages of 1995.

Kramer was there in 1995, writing HTML by hand, building web servers in C, as convinced as anyone that the web page was the future. "I was as excited about HTML in 1995 as I am about agents today," he says. That enthusiasm was warranted. The web page was real and important. It's also still with us. But nobody in 2025 talks about web pages the way they talked about them in 1995. The conversation moved to web applications, to native mobile, to the cloud. The page was a primitive — a necessary stepping stone — not the destination.

That framing reorients the entire current conversation about agents. The agent-native design questions that builders are wrestling with right now — how to structure tool calls, how to manage context, how to handle multi-model orchestration — are real engineering problems worth solving. But Kramer's point is that solving them is table stakes, not the endgame.

Each decade had a job

To build the case, Kramer runs through computing history decade by decade, and the structure of his argument is worth following carefully because it's doing real analytical work, not just providing color.

The 1950s were about getting human intent into machines — programming languages, compilers, the first attempts to bridge mathematical formalism and something resembling natural language. The 1960s were about interactivity: Ivan Sutherland's Sketchpad, time-sharing systems, the idea that a machine could have a two-way dialogue with a person. The 1970s scaled the abstractions — relational databases, declarative languages, Smalltalk and object-oriented programming. All of this built the substrate for the personal computer of the 1980s.

The pattern Kramer is tracing: each era's big contribution was not the final product but the primitive that made the next era possible.

VisiCalc is the example he returns to most, and it's well-chosen. The first spreadsheet didn't put accountants out of work. It made vastly more accounting possible and generated categories of work that couldn't have been described when doing a screenful of calculations required a room full of people. Kramer offers this as a reply to the mass-unemployment concern that surrounds current AI development — not a dismissal of the concern, but a historical data point that complicates the straightforward displacement story. Whether the analogy holds at the scale and speed of current AI capability is a question he doesn't fully resolve, and it's the right question to leave open.

The multimodal thread

One of the more interesting moves in Kramer's argument is his insistence that multimodality is not a new feature of AI systems — it's a thread that runs through the entire history of the web. "The single most important thing about the web," he says, "was that it was multimodal from the very beginning. More even than the GUIs of the 1980s, the web anticipated that text and audio and video and data were not different things to be used in different programs. They belonged together."

That framing makes the current excitement about multimodal AI models look less like a breakthrough and more like a homecoming — the infrastructure finally catching up to an idea that was always implicit in how the web was conceived.

Apple's 1987 Knowledge Navigator concept video gets substantial treatment here because it sits at almost the exact midpoint of Kramer's timeline and predicted, with uncomfortable accuracy, what we now call an AI agent: a foldable tablet with a conversational voice assistant, access to personal and global information, real-time video, autonomous task delegation, and something resembling continual learning. In 1987, this was a fantasy. Today it is a product description.

Kramer notes that Tavus — identified by Kramer in his talk — recently released a reimagined Knowledge Navigator video built entirely on current, available technology. Four minutes, one take, no special effects. The comparison is striking not because the technology is magic but because the gap between concept and implementation, which looked unbridgeable in 1987, has closed.

Where Nadella fits

Satya Nadella, speaking recently on a crossover episode of the No Priors and Latent Space podcasts, described the current state of AI product development in terms that Kramer uses as a signpost for where the field actually is. Here is how Nadella framed it — and notice the specificity of the language:

"You kind of want the harness to define the models, the data, and the tools so that you have a loop across those three... all of them are multimodel harnesses with tools access so that you can do this progressive disclosure of tools, even so that they're token efficient, and then you're feeding it with very rich context."

"Multimodel harnesses." "Progressive disclosure of tools." "Rich context." These are the engineering concerns of right now. Kramer's point is that they are the concerns of the web-page era — the moment before the architecture settles into something more durable.

The term he uses for the current frontier is "agents plus": multi-model harnesses, software co-pilots embedded in every application, organization-level orchestration. It's a more precise description of where the industry is heading in the next 18 to 24 months than most of what gets published under the AI banner. Reaching the point where agents reshape coding at scale is itself a step in this longer arc, not the destination Kramer has in mind.

Gradient Bang as a test bed

The talk closes on a project called Gradient Bang — a massively multiplayer game that Kramer and collaborators built as an explicit exercise in the primitives they believe come next. The game runs hundreds of inference calls simultaneously, with LLMs at the core of every interaction. Kramer says it couldn't have been built a year ago.

What makes it interesting analytically is not the game itself but what it was designed to stress-test: asynchronous non-blocking context compression, long-running subagents that share context, progressive skills loading, dynamic interface generation, and conversational voice. These are architectural patterns, not product features. Kramer is using a game to figure out what the scaffolding for AI-native software needs to look like — the same way early web developers used increasingly complex pages to discover that what they actually needed was an application framework.

The analogy to 1995 is doing a lot of work in Kramer's argument, and it's worth asking how far it extends. The web had network effects and an open standard at its core. AI-native software is being built on top of proprietary models with pricing structures that can change overnight and capability curves that are genuinely hard to predict. The path from web page to web application took roughly a decade and benefited from a relatively stable infrastructure layer. Whether the path from agent to AI-native software follows the same arc — or compresses, or forks — is not something Kramer claims to know.

What he does claim is that the people building agents right now are in the position of people writing HTML in 1995: doing important, real work on something that will turn out to be a primitive for whatever comes next. The reasonable response to that, if you believe it, is not to stop building agents. It's to build them while keeping one eye on the scaffolding you're actually constructing.

Bush got the destination right in 1945. The computing age gave us eighty years of road. The question now is whether the intelligence age compresses that, and what we miss if it does.

---

Bob Reynolds is Senior Technology Correspondent at Buzzrag.