Good Architecture Deletes the Signals Your Agent Depends On

AI coding agents rely on similarity-based retrieval to understand the impact of a code change, but this approach is flawed. The code that resembles a component is often not the same as the code that breaks when it changes, which is the true dependency. Good software architecture, which uses abstractions and contracts, intentionally removes the textual signals that similarity search depends on. This creates a gap where agents cannot see the true dependencies, especially across different repositories or services. The solution is not better search but a better understanding of the system's structural dependency graph, which is often computed at build time rather than being explicitly present in the source code.