The Biology of Software

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI describes AI systems gaining capability through composition and augmentation—giving a model tools, letting it call external systems, chaining it into agentic workflows, combining it with retrieval over a knowledge base. Kay's biological metaphor reveals what this convergence is: not an engineering kludge but the rediscovery, by a different route, of how nature organizes intelligence. Each capability gain from composition rather than from raw scale alone mirrors the mechanism Lynn Margulis demonstrated in the cell: a host acquiring a competence it could not have evolved independently, by incorporating an organism that had already evolved it. The AI system that uses a calculator, a code interpreter, a search engine, and a memory is not simply a bigger model. It is an ecological system of autonomous agents each contributing a competence, integrated by a routing mechanism into a whole more capable than any part.

Kay's biological metaphor also reframes the AI interaction itself. In the mechanical model—input, process, output—the user is a client and the model is a tool. In the biological model—message, response, message—both are autonomous participants in a conversation, each processing according to its own internal logic, contributing to an emergent behavior that neither designs or fully controls. The quality of the interaction depends not on the model's capability alone but on the quality of the dialogue—on both participants contributing something the other cannot, and integrating their contributions into something richer than either could produce alone. This is the design ideal the cycle aspires toward: human and AI as genuine partners in a medium, rather than user and tool in a transaction.

The specific failure Kay identifies is that the biological vision has been adopted at the architectural level (multi-agent systems, mixtures of experts, compositional intelligence) while remaining unimplemented at the interaction level. The routing networks and mixture-of-experts gates are structured as Kay described. But the interaction between the human and the composite system is still designed as command and execution: the human specifies, the system delivers. The autonomy that Kay wanted each object to contribute—the capacity to respond not with what was requested but with what the situation requires, to challenge assumptions and redirect attention—is not a feature of current AI tool design. It is the design that remains to be built.

Origin

The biological metaphor emerged from Kay's intellectual formation before he was a computer scientist. He had studied molecular biology at the University of Colorado, and his understanding of how cells work—as self-contained units communicating through chemical messages, with organism-level behavior emerging from cell-level interactions—gave him a model of organized complexity that the mechanical paradigm could not provide. When he encountered Ivan Sutherland's Sketchpad at Utah and the programming language Simula, he saw in both the possibility of software organized as entities with local state and behavior, and the biological metaphor crystallized: software should be a community of autonomous objects, communicating by message-passing, with emergent behavior arising from their interactions.

The result was Smalltalk, developed at PARC through the 1970s, which gave object-oriented programming its biological soul. Kay was explicit about the biological inspiration, drawing diagrams that looked more like cell biology than circuit diagrams, and insisting that what he meant by “object-oriented” was the ecological insight of autonomous agents communicating through messages—not the class-hierarchy syntax that the industry eventually adopted under the same name. The industry took the vocabulary of classes and objects while discarding the ecological idea: the objects in mainstream Java or C++ are data structures with attached procedures, organized into corporate hierarchies by programmers still thinking mechanically about control. The biological metaphor was replaced by an organizational metaphor. What most programmers call object-oriented programming is not, in Kay's assessment, what he meant.

The convergence between Kay's biological vision and the architecture of multi-agent AI is one of the more striking vindications in the history of computing. Kay designed the ecological architecture for pedagogical and aesthetic reasons—because it was the right model for how intelligent systems should be organized—and the AI field reached the same architecture through empirical pressure, because the monolith ran out of room and composition proved more capable. The two routes arrived at the same destination, which is some evidence that the destination was correct.

Key Ideas

Objects as autonomous agents. In Kay's biological model, a software object is not a data structure with procedures attached. It is an autonomous agent that carries its own state, follows its own logic, and communicates with other objects by sending and receiving messages. The object does not know the internal workings of the objects it messages; it knows only the interface. Complex system behavior emerges from the interactions of these autonomous agents, the way organism behavior emerges from cell interactions. This is structurally identical to how large language model agents in a multi-agent system operate: each model receives messages (prompts), processes them according to its own parameters, and returns messages (responses), with the system's capability emerging from their composition.

Late binding. Kay's related principle: deferring decisions until the latest possible moment, keeping the system open, flexible, responsive to new information. In a late-binding system, behavior emerges from ongoing dialogue between components rather than being specified in advance. Biological systems are maximally late-binding: cells do not commit to fixed plans but respond to messages from their environment, adapting dynamically. The large language model is the most late-binding system in computing history, deferring all decisions to the moment of interaction, responding to each message in context. Late binding enables the exploration and discovery that the mechanical paradigm forecloses.

Emergence over control. The designer of a biological system does not specify the system's behavior. She designs the objects—their capabilities, their message interfaces, their local logic—and the system's behavior emerges from their interactions. This requires a different relationship to design: not engineering a machine but cultivating an ecology. The shift is from control to coordination, from mechanism to ecosystem, from prescribing outcomes to setting conditions under which productive outcomes can emerge. The multi-agent AI system whose capabilities exceed the capabilities of any individual component is living this principle, whether or not its designers are thinking in Kay's terms.

The unimplemented ideal. Kay's biological vision has been adopted at the architectural level of AI systems but not at the interaction level. The routing networks, mixture-of-experts gates, and multi-agent compositions are structured as Kay described. But the individual AI agent—in its interaction with the human user—is not designed as an autonomous biological object that contributes its own perspective and challenges the human's assumptions. It is designed as an obedient tool that delivers what is requested. The autonomy that gives the biological metaphor its power remains an unrealized design aspiration at the level of human-AI interaction.