AI Agents Era: 'Understanding Code' Still Matters
Original: Understanding is the new bottleneck
Why This Matters
Defines a new human role in AI-driven development: creative collaborator, not just verifier.
Geoffrey Litt argues that as AI agents write more code, human understanding remains essential — not merely for verification, but for active creative participation. He presented three techniques at the AI Engineer conference in July 2026 to help developers efficiently comprehend agent-generated code.
Geoffrey Litt, speaking at the AI Engineer conference in July 2026, challenged the notion that developers can simply step back and let AI agents handle everything. His central argument: understanding code is still critical, but not primarily for verification purposes. As agents become better at verifying their own work, the human role shifts from checking correctness to active creative participation across the many iterative loops of a project. Litt warns of 'cognitive debt' — a concept popularized by Margaret Storey and Simon Willison — where failing to understand what agents are building accumulates into a long-term liability, similar to technical debt. To address this, Litt draws inspiration from education and proposes three practical techniques: (1) Code explainer docs — structured HTML, Markdown, or Notion documents generated via a custom '/explain-diff' skill that summarizes what changed and why; (2) Quizzes to actively test one's own understanding of a system; and (3) Micro-worlds — interactive environments that let developers play with and internalize how a system behaves. Litt emphasizes that a rich conceptual model of the codebase is what enables developers to think fluently and creatively about how to evolve a product.