Why the lean AI coding harness is failing

augment code is betting that pre-indexing repositories via embeddings and vector databases provides better context for ai agents than the real-time grep methods used by tools like claude code. vp of engineering vinay perneti argues that semantic retrieval allows agents to access conceptually relevant code more effectively than simple text matching. this challenges the 'lean harness' philosophy favored by anthropic, which avoids building structured context in advance to avoid obsolescence. as models improve, the debate over how to manage context windows remains central to ai-assisted development workflows.