Complexity as the Agent's Worst Enemy

Building Pi, and what makes self-modifying software so fascinating

Topic 9 of 14

Published Apr 29, 2026 · Clip 1:00:22–1:02:45 (2:23)

Mario explains why complexity is the agent's worst enemy: with a ~200k token context window, an agent can see only a fraction of a large codebase, and information retrieval for relevant code remains unsolved. Agents generating huge codebases eventually can't even ingest their own output. He also argues models converge toward the mean of internet training data — mostly garbage and cargo-culting rather than excellent projects like Linux.