The prompt before the architecture
Third article in the series on building Bloom in a week: why most of the perceived rigidity in a multi-agent agent comes from the prompt layer, not the architecture, and how to hard-code behavioral safety invariants.