AI Coding Stopped Being About Writing Code

Simon Willison shipped a release candidate of sqlite-utils that was mostly written by an AI agent for about $149. Software is now as cheap as dinner. It's the wrong number to fixate on. Before shipping, Willison had it do a final review and it surfaced five release blockers he hadn't hit himself, including a delete_where() bug that never committed its transaction, found in review, not in generation.

The generation was cheap and mostly fine. The value was in the verification pass. That's not a code-generation win. It's a code-understanding win. This is exactly the split we keep arguing about at Navera: modernization, and increasingly all serious AI-assisted engineering, is a code-understanding problem before it is a code-generation problem. Writing the new thing is the easy half. Reliably knowing what the old thing does — and whether the new thing still does it — is the hard half, and it's where the money and the risk actually live.

Neo4j spent the week hammering the same nail from the data direction, laying out why agentic RAG needs a graph rather than vector similarity alone. Their argument is the one we live by: vector search returns text that looks similar, not context that shows how things connect, For a legacy codebase, "how things connect" is the entire game — call graphs, data lineage, the dependency you didn't know existed until it broke in production.

So here's the through-line for anyone modernizing legacy code. The week's cheap-code milestone is real but beside the point. Trust doesn't come from the model. It comes from the verification layer you build around it — and that layer is the product. Contact Narvea Ventures to know more on how we solve this!