The Complexity Ratchet in the Age of AI Coding: Why 90% Test Coverage Matters

Garry Tan argues that while working on large-scale projects with AI agents, the traditional software engineering 'trade-off between speed and quality' has broken down. He presents the Complexity Ratchet as the core mechanism behind this.

How the Complexity Ratchet Works The ratchet is a structure that ensures code quality only moves forward, never backward. With every work session with an agent, the following three elements accumulate:

• Tests: Define what the correct result is
• Documentation: Record the reasoning behind decisions
• Eval results: Establish a baseline for quality

Because AI agents load these accumulated assets as context for the next task, they become unable to perform work that breaks existing tests or degrades quality.

How AI Has Transformed Software Engineering

• A shift in the error model: In the past, complex processes were built to prevent errors, but now the paradigm is shifting toward agents diagnosing and fixing errors instantly.
• Expansion of the complexity ceiling: The upper limit of system complexity has expanded from human memory capacity to the size of the context window an agent can handle.
• Realizing 90% test coverage: By having AI agents tirelessly handle the last 20% of edge-case testing—which was previously too costly for humans—organizations can now routinely achieve aviation/medical-grade levels of advanced verification.

In conclusion, the true value of AI coding lies not merely in its speed, but in enabling extremely high levels of verification that were previously abandoned due to cost constraints, thereby raising the complexity ceiling of software that developers can handle.