What Happens When the Model Eats the Stack? Rethinking the Research Agenda for Data Agents to Withstand the Bitter Lesson

A recent paper argues that as LLMs improve, they absorb hand-engineered data-agent scaffolding, echoing Sutton's Bitter Lesson: newer models (GPT-5.6 Sol vs o3) beat human-designed agents like Agentar-Scale-SQL on both accuracy and efficiency, with turns per query dropping from 23.2 to 6.0. Schema exploration and semantic mistakes (idiosyncratic business definitions, join keys, undocumented schemas) become the dominant remaining failure mode. The paper proposes building a persistent, database-stored semantic context layer to avoid redundant exploration, though its technical proposal is critiqued as superficial. The commentary also pushes back on the paper's implicit rebuttal of a prior 'agentic speculation' paper, invoking Jevons paradox, and proposes TLA+ as a formal foundation for maintaining consistent semantic context and catching invariant violations before query execution.

Nguồn: https://muratbuffalo.blogspot.com/2026/09/what-happens-when-model-eats-stack.html. 8 Sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.