Agents are the web pages of the current software era, and AI native applications will come after them.
2
Each major computing era introduced new primitives, from programming languages and interactive systems to databases, personal computers, the web, mobile devices, and cloud infrastructure.
3
Gradient Bang tests AI native software patterns such as shared-context subagents, asynchronous context compression, progressive skills loading, dynamic interfaces, and conversational voice.
Summary
Kwindla Kramer places current agent development in a longer history of computing. Vannevar Bush's 1945 essay, As We May Think, anticipated technologies including OCR, speech interfaces, hypertext, search, and wearable cameras. Kramer compares today's agents with the web pages of 1995, which were useful but later gave way to full web and mobile applications. He traces the primitives that made earlier eras possible: programming languages, interactive computing, scalable abstractions, spreadsheets, networks, mobile devices, and cloud infrastructure. VisiCalc is his example of software expanding the amount and type of work people can do rather than simply removing a profession. He argues that current agents and multimodal harnesses will lead to AI native software built around new interaction patterns. His examples include a real technology remake of Apple's Knowledge Navigator and Gradient Bang, a multiplayer game with language models at the center of every interaction.
Vannevar Bush imagined much of the modern computing stack before it existed
Kramer begins with Vannevar Bush's 1945 essay, As We May Think. Bush described document display, scanning, OCR, speech-to-text, text-to-speech, programming languages, hypertext, search engines, data networks, wearable cameras, voice interfaces, and brain-computer interfaces. Kramer sees the essay as a useful comparison for the current intelligence age. Bush wrote at the beginning of digital computing, while today's builders are working at the beginning of an era centered on machine intelligence. The question is what software primitives will emerge from this period.
Kramer was a programmer in 1995, when web pages occupied the same space in developers' thinking that agents occupy now. He wrote HTML by hand, built web servers and indexing software in C, and made authoring and management tools in Perl. Web pages remain useful, but most software work moved toward web applications and native mobile applications. Kramer expects a similar transition from agents and multimodel harnesses to fully AI native software. Agents are an important starting layer, rather than the final form of the applications people will build.
Each computing era solved a different problem between machines and people
Kramer's timeline moves from the abacus to stored-program computers, personal computers, and the current agents era. In the 1950s, programming languages and compilers made it easier to transmit human intent to machines by combining mathematical formalisms with forms closer to natural language. The 1960s focused on interaction and two-way dialogue, with Ivan Sutherland's Sketchpad providing an early example of graphical programming. In the 1970s, relational databases, declarative languages, Smalltalk, and object-oriented programming provided abstractions for larger data sets and more powerful hardware.
VisiCalc expanded accounting work instead of eliminating accountants
Kramer uses VisiCalc as an example of a new computational abstraction becoming broadly useful. The spreadsheet made numerical computing interactive and two-dimensional, and its descendants include Google Sheets and Microsoft Excel. He connects this history to fears that AI will cause mass unemployment. VisiCalc did not put accountants out of business. It made much more accounting-like work possible and enabled categories of work that were difficult to imagine when a screen of calculations required a roomful of people. For Kramer, this is a model for how new software capabilities can expand work.
The web's defining feature was combining media in one networked environment
Kramer argues that the web was multimodal from the beginning. Text, audio, video, and data were brought together instead of being confined to separate programs. He sees the web as a practical attempt to realize ideas similar to Apple's 1987 Knowledge Navigator concept. The following decade focused on making this multimodal, networked computer mobile and continuously connected. Cheap screens, ubiquitous cameras, and increasingly capable devices also made it possible to imagine future interfaces that were almost, but not quite, buildable.
AI native software can now make older interface visions work
Kramer describes a progression from calculators to computers, personal computers, and the global cloud computer. He argues that current infrastructure makes it possible to build working versions of ideas shown in the Memex, Iron Man's Jarvis, and Apple's Knowledge Navigator. A team led by Tavis released a new Knowledge Navigator video made entirely with real, available technology. Kramer says the four-minute video is one continuous take. It feels familiar because it follows the 1987 concept, while also being new because the voice assistant, calendar interaction, and surrounding system are real.
Gradient Bang uses a game to exercise emerging AI software primitives
Gradient Bang is a massively multiplayer game built with language models at the center of every interaction. Kramer says hundreds of inference calls are active at any moment, something the team could not have built even a year earlier. The project tests asynchronous, non-blocking context compression; long-running subagents that share context; progressive skills loading; dynamic interface generation; and conversational voice. In the demonstration, agents manage trade and exploration loops, record future actions, answer questions about ships, display task history, hide the map, and handle a voice conversation about selling a ship.
"The web page is still with us and it's still important and useful. But today we talk a lot more about web applications and native mobile applications than we talk about web pages."04:25
Who should watch
You are building agents and want a historical model for what may come after the current agent stack.
Your product depends on voice, multimodal interaction, shared context, or many concurrent model calls.
You are thinking about AI's effect on work and want Kramer's VisiCalc-based argument rather than a simple automation forecast.
