The "predefined patterns reduce token costs" framing resonates. We saw the same thing building empla.io. Every time we left a backend decision open to the agent, it burned 3x to 4x the tokens exploring before committing. The counterintuitive part is that declarative query languages save more tokens than they save for humans, because the agent does not have to plan imperative control flow step by step. Two open questions for the Instant team. How are you handling schema evolution when an agent decides mid-session that it needs a new relation? That is exactly where Supabase falls over for us. And do you expose a cost budget primitive per session, or does the user have to instrument that separately?