Episode Details

After an exhaustive survey in July 2026, the answer is clear: nobody has built a software development protocol designed from the ground up for AI agents. Not for spec definition, task decomposition, execution tracking, or agent-to-agent handoff. Markdown is still the state of practice, and agents routinely misparse task state from checkbox characters. This episode walks through the five candidate primitives for such a protocol, debates which one is actually load-bearing, and wrestles with the uncomfortable possibility that the artifact has to stay human-legible because a human has to remain accountable for it.