Conway's Law and Programming Languages

Casey sparked my interest in Conway's Law, which explores the relationship between information flow among designers and system structure. The paper posits that interfaces, such as APIs and documentation, form where communication crosses boundaries among humans or teams, reflecting the org chart. When I apply this model to agents, I find it doesn't translate neatly, as agents operate at a much faster pace than humans but lack shared prior knowledge with users.

Agents can only learn about their environment through tool calls and data in their context window, limiting the scaling of teams even with shared priors and maximum bandwidth. This leads to context pollution, particularly in multi-agent systems, resulting in suboptimal performance compared to dedicated single agents. Like human teams, agents suffer from working in silos, and interfaces in agent-agent or human-agent interactions are dynamic bounded channels with capacity limits.

The speed at which humans read agent outputs and the decisions resulting from understanding the output become critical bottlenecks. While humans and agents share a common bottleneck in processing and decision-making, the choice of programming language for agents matters more for humans than for the agents themselves. Code readability and expressiveness are crucial for reducing communication tax, and languages like Clojure offer reduced complexity by eliminating classes of bugs through immutable data.

However, other languages may also provide similar benefits, allowing humans to focus on higher-level problem-solving and improving the efficiency of communication between humans and agents.