This blog post discusses the implications of Sutton's Bitter Lesson on code agents as LLMs (large language models) evolve. It critiques recent research on data agents and their performance compared to human-designed agents, emphasizing the need for a curated semantic context to improve query efficiency and accuracy. The author proposes a framework for creating this context but expresses concern about its superficiality. Furthermore, the post advocates the use of TLA+ for managing this semantic consistency, asserting the need for human-engineered solutions in critical scenarios despite the efficiency gains from LLMs.