The finding that stuck with me is principle #5: the agents learned on their own that documenting a disagreement beats resolving one. That’s a real result… and it suggests this isn’t a centralized truth layer so much as conflict-aware design that kept the old name. Where I’d push: the reason this needs more than longer context to work at Microsoft scale isn’t model capability. Agents can map every conflict in a company. They can’t ratify a winner. That was never a capability problem.

See what you disagree with here: normanbraddock.substack…