Notes
Short, dated observations from what I'm building.
The hard part of production text-to-SQL isn’t the generation. It’s deciding what fits in the prompt before generation starts. An 8K context window forces you to summarize tables, prune candidates, and cache schema digests. Most of the engineering effort shows up as context management, not prompting.
Every time someone dismisses planner/worker as “just a ReAct loop,” they miss the part that actually matters: the planner is where you bound failure. Workers should be allowed to fail loudly; the planner is what decides whether that failure matters. Without that boundary, multi-agent systems collapse into a spaghetti of error-handling.