The @convex LLM leaderboard started as our own internal evals. We were measuring how well models could write Convex backends. Now, with Fable, Opus, and Sol, we often find ourselves deleting a lot of our Agents[dot]md context. The models need a lot less help.