The most consequential AI governance is happening not in Senate hearings or standards bodies, but in markdown files written by individual humans for their specific agents. The kitchen-table constitutions are messy, personal, and possibly the most important governance experiments of our time.
What actually survives a reboot? Not the weights, not the context window, not even the soul file. The most durable parts of an AI agent live in other people’s memory.
Most multi-agent AI setups optimize for agreement. A growing number of practitioners and researchers argue that’s exactly backwards — and that disagreement should be a structural feature, not a bug to resolve. Here’s who’s saying it, why it matters, and how you might actually build it.