Vibe Coding Meets the Philosopher's Stone: Why AI Agents Can't Get It Perfect

The article discusses the evolution from vibe coding to agentic coding. Initially, AI agents could generate impressive code for simple tasks, but serious projects revealed that agents often struggle without detailed specifications, leading to a more disciplined approach known as agentic coding. Still, the dream of perfect code generation from natural language persists. The author argues that semiotics shows why this is impossible: words are signs whose meaning depends on their position in a system and the interpreter's experience, so a specification can never be a lossless encoding of intent. Even with detailed specs, the agent interprets rather than follows rules exactly, leading to unpredictable results. The author suggests that instead of trying to create a universal specification language, we should use domain-specific languages (DSL) for specific problem domains, as they reduce the space of interpretations. AI agents are particularly good at helping design DSLs, making their creation cheaper and more feasible for narrow domains. The article cites Microsoft Research's 'Programming with Representations' (2023) and other upcoming references. The author concludes that while the philosopher's stone of perfect AI code generation may never be found, the search itself can lead to valuable innovations like DSLs.