Why AI Agents Cannot Change Software Systems

While LLMs can generate correct code, they cannot safely modify real software systems because they lack knowledge of system invariants, constraints, and interdependencies that are critical to maintaining production stability. The author argues that transformative changes—which alter running systems—require understanding operational patterns and integration boundaries that AI models cannot perceive or reason about, distinguishing this from additive tasks like reading and planning. Despite progress in agentic AI systems, current implementations remain unsafe for production use without significant human oversight, contrary to claims that "fully automated agents are here."