Stanford-MIT: AI agents consume millions of tokens per coding task
A joint Stanford-MIT study found AI agents consume millions of tokens per coding task, exposing dramatic inefficiencies in current models. The research is a concrete cost and scalability warning for organizations building agentic coding loops in production.