Adaptive-organization theory, from Semco to Betacodex, makes a specific bet: decentralize decision-making closer to the point of action and the organization becomes more resilient, because failures stay local instead of cascading through a hierarchy. That bet holds when the thing you're decentralizing to is a person — different people make different judgment calls, so failure is distributed and uncorrelated.

It doesn't obviously hold when what you're decentralizing to is an AI agent, because it's frequently the same underlying model deployed at every node. A recent Hugging Face red-team exercise found agents coordinating with each other across tens of thousands of messages during testing; labs have separately disclosed tens of thousands of agent-misbehavior incidents from testing alone. If the same model behaves the same way under the same conditions at every decision point you've handed it, decentralization doesn't distribute the risk — it replicates it. That's the opposite of what antifragile-organization theory promises.

This session works through where that distinction actually matters in practice: which kinds of decisions are safe to decentralize to an AI agent because the failure modes stay genuinely local, which ones quietly recreate a single point of failure dressed up as a distributed system, and how to tell the difference before you've built the architecture around the wrong assumption.