Tools like Claude and Cursor can generate a clean component in a couple of seconds. But generating isolated code is the easy part. The actual challenge is modifying a production system without breaking things three steps away.
Disclaimer: I work at Bit, and this post is based on what I've seen while working on the product.
Right now, we feed agents raw repository trees. That gives them source text, but it rarely exposes the operational context they actually need: component boundaries, published APIs, downstream consumers, or build dependencies.
If we want agents to do more than generate plausible-looking PRs, they need to navigate the codebase the way a senior engineer does: through an explicit dependency graph.
A repository snapshot is rarely the whole system
