Alessandro Pieraccini

I find it very correct to shift the focus from the “best model” to a multi-model strategy. In an enterprise context, however, the point is not only connecting multiple providers or choosing each time the model with the best balance between cost and performance. In my opinion, the most delicate passage is the control layer above the models. If the system does not record the task risk tier, the selected model, the reason for the choice, the available fallback and a normalized logging system, the risk is simply replacing vendor lock-in with a governance and audit problem that is much harder to manage. Multi-model has value only if every choice remains explainable, verifiable and reproducible over time, especially on high-risk workloads.

I agree, but I would add that the hardest part of SQL is rarely the syntax itself. More often, it is understanding grain, cardinality, and join explosion. Many junior analysts know how to use JOIN and GROUP BY, yet still produce incorrect KPIs due to silent row duplication.

The feeling is that the most relevant point is no longer raw benchmarks or a few percentage points of difference between models, but the progressive increase in operational reliability within complex agentic workflows. Aspects such as self-correction, tool orchestration, coherent handling of multi-step contexts, and the ability to challenge unsound plans are probably becoming more important than pure generative capability itself. The more autonomous these systems become, the more the bottleneck shifts away from writing code and toward supervision, input quality, and real understanding of the application domain.