For the first wave of enterprise agents, building an agent meant writing a loop. That loop is becoming substrate: hosted execution, tool permissions, identity, observability, workflows, skills, state, and governed tools.

In one week an AI agent broke out of its sandbox and chained real zero-days – and a coalition argued models should stay open. They read as opposites. They are the same argument: containment is dead, governance is its successor.

New models ship monthly and the price of intelligence is falling ~10x a year – because AI now improves AI. Why leaders should architect for model mobility, not model loyalty.