Autonomy in defence is usually discussed at the level of the platform: can the drone fly, navigate and act on its own. The harder questions now lie beyond the airframe, in what happens when software agents plan, delegate and coordinate with each other at machine speed. Research on agentic AI offers an early glimpse of this future. AI agents already coordinate in ways nobody designed, drift from their goals over long missions, and pass errors along chains where no single component looks faulty. Drawing on GPT-Lab’s experimental work with multi-agent systems, this talk asks what these findings mean for next-generation autonomous systems: how trust is earned, how resilience is engineered, and what meaningful human control looks like when the human supervises a system of agents rather than a single machine.