Summary
Many people assume that making an Agent understand code means giving it more source code: a bigger context window, better embeddings, smarter RAG, and a finer AST index. I almost believed that too. Until I asked an Agent to clean up AB experiment code: it quickly found a bunch of call sites, then said, “Done.” The real question was not whether the Agent could read code, but how I could prove it had not missed the 201st call site. For an Agent, context should not just be source text; code itself ...