The Day After

The developer did everything right. The agent did everything right. The system failed them both. That's not an edge case or an early-adopter growing pain. It's the pattern showing up on every team that has pointed an AI agent at a real codebase and waited for a productivity gain that never fully arrived. The agent generated code. The code compiled. The code passed the linter. Then it failed in staging because the constraint it broke didn't exist anywhere in the repository. It lived in the memory of a developer who left, in a Slack thread from three years back, in a decision that got made in a meeting and never written down.

Most organizations are treating this as a tooling problem. They're adjusting prompts, switching agents, widening context windows, hiring prompt engineers. Chapter 1 argues the diagnosis is wrong. The problem isn't the agent. It's the system the agent has been asked to work inside, a system that was never designed to be navigated by anything that couldn't just ask a person for help. Humans have always been able to compensate for a confusing codebase by talking to each other. Agents can't do that. So every gap in what the system makes legible, every piece of knowledge that never got written down, becomes something an agent runs straight into.

This is the chapter that names the shift and lays out the competitive map forming around it. Some companies are already pulling ahead because they understood what actually changed. Others are still tuning their prompts, solving a problem that isn't the one in front of them. What separates those two groups is the subject of everything that follows.