AI Agents: The New Primitives of Software

Kwindla Hultman Kramer, from Daily, delivered a compelling keynote at the AI Engineer World's Fair, charting the historical progression of digital computing and positioning the current AI agent era within that lineage. Kramer, whose company Daily develops developer infrastructure for real-time audio, video, and AI, and is the team behind the widely used Pipecat framework, drew parallels between past technological shifts and the burgeoning field of AI agents.

From Vannevar Bush to AI Agents

Kramer began by referencing Vannevar Bush's seminal 1945 essay, 'As We May Think.' Bush, an engineer, academic, and civil servant, possessed a deep understanding of various technologies and, in his essay, foresaw numerous advancements, including document display on screens, OCR, speech-to-text, text-to-speech, programming languages, hypertext, search engines, data networks, and even early concepts of brain-computer interfaces.

Kramer highlighted the significance of Bush's predictions, noting that we are now living through a similar transformative period, which he termed the 'intelligence age.' He stated, "I've been thinking a lot about 'As We May Think' lately because Bush wrote this essay right at the very beginning of the computing age. And I think it feels to most of us like we're working right at the beginning of a new age, the intelligence age."

The Evolution of Software Primitives

The presentation then traced the evolution of computing primitives. Kramer compared the current focus on building AI agents and 'agents plus' (multi-model harnesses and embedded software co-pilots) to the early days of the World Wide Web. He recalled spending time writing HTML by hand and building web server software in C in 1995, a period when 'web pages' were the primary focus, much like 'agents' are today.

Just as the web evolved from static web pages to dynamic web and native mobile applications, Kramer posited that we are now charting a path towards entirely AI-native software that will supersede the current agent paradigm. He outlined a historical timeline, starting from the abacus, moving through stored-program computers and personal computers, and arriving at the current AI agents era, projecting a future that builds upon these advancements.

Key Technological Leaps

Kramer detailed the key developments that enabled these transitions:

• 1950s: The development of the first programming languages and compilers, bridging mathematical formalisms with natural language.
• 1960s: The push for interactive systems and two-way human-computer dialogue, alongside graphical programming systems like Sketchpad.
• 1970s: The creation of abstractions that could scale, such as relational databases and declarative languages, alongside advancements in programming like object-oriented programming.
• 1980s: The personal computer revolution, exemplified by the Macintosh and Windows, which brought tangible benefits like VisiCalc, making complex tasks accessible to a wider audience.
• 1990s: The rise of the network and the World Wide Web, characterized by its multimodal nature, integrating text, audio, video, and data.
• 2000s: The focus on mobility and continuous connectivity, putting supercomputers in people's pockets.
• 2010s: The development of cloud infrastructure, laying the groundwork for AI training and inference.

The Future: AI-Native Software

Kramer emphasized that the technologies predicted by Vannevar Bush, such as the Memex, are now becoming a reality. He referenced a recent reimagining of the 'Knowledge Navigator' video, built with current technology, showcasing a foldable tablet, conversational AI with personality, real-time video and computer vision, and seamless integration of these capabilities.

He concluded by discussing a multiplayer game project, 'Gradient Bang,' built from the ground up with LLMs at its core, demonstrating real-time agent orchestration, asynchronous sub-agents, and dynamic UI generation. This project, he noted, would not have been possible even a year ago.

Kramer expressed his excitement for building not only current AI agents but also the 'next, next thing' that will emerge from this AI-native software revolution.