Ask a coding agent to change a component and it will do it. Ask it what that change broke and it will guess. The author, co-founder of Bit Cloud, explains that code-retrieval systems answer the wrong question: they find code that resembles the query, but an agent needs code that breaks when something changes.

Those two sets overlap far less than tooling assumes. Cursor, Claude Code, and Copilot all sit in front of the same gap because dependency information lives outside any repository they can see. DORA's 2025 State of AI-assisted Software Development report confirms this, showing time saved during generation being re-spent on auditing, with around 30% of developers reporting little or no trust in AI-written code. The author uses a concrete example: a learning platform with a mentor chat widget installed as a separate published component. The page importing it exists in another repository, but working from the widget, nothing indicates that page exists.

The article argues that architecture deliberately removes vocabulary overlaps that similarity indexes rely on, making resemblance a poor proxy for dependency.