What AI Agents Speed Up in PhD Research, and What They Can't

Large language models are significantly accelerating the execution layer of AI research, including tasks like coding, running experiments, and drafting papers. However, these tools cannot replace the deep domain knowledge required to understand the context and limitations of data, as they lack the context to avoid fundamental errors. This shift will make researchers more like team leads who direct AI tools, making domain expertise the primary advantage for a research lab. Consequently, AI labs are predicted to become more interdisciplinary, while human qualities like curiosity, problem selection, and the ability to build trust remain irreplaceable.