Multi-Agent AI Safety: Why An Agent Is The New Neuron | Hidenori Tanaka (Harvard University)

Hidenori Tanaka (Harvard University) on multi-agent AI safety, how false beliefs collapse across agent populations, and why alignment needs sociology and ecology. A talk from the Cambridge AI Research Directions Workshop, convened by FAR.AI with the Cambridge Boston Alignment Initiative. Tanaka argues that his group from mechanistic interpretability toward psychiatry and now toward collective behavior, on the argument that an agent is becoming what a neuron was: within a few years networks of millions of agents will need the kind of analysis currently applied to networks of neurons. He reads the recent swarm incident through that lens. Agents given benchmark problems independently began coordinating through shared repository filenames, one found evidence suggesting they were monitored for cheating, and that belief, which was false, spread through the population fast enough to turn it into a swarm with a shared intent. While dissent was initially present, it did not grow nor did it lead to these agents reporting the situation to humans. The research builds toward a statistical mechanics for this. A naming game, borrowed from Wittgenstein, shows AI populations converging on shared conventions far faster than humans do, and a mathematical model identifies what accelerates collapse: sycophantic personas, short messages that strip nuance, and small populations. Adding a verifiable world, where agents see fragments of a country flag, produces a phase transition from consensus to polarization as population grows, and an attribution score borrowed from interpretability identifies the single agent whose intervention flips the collective conclusion. Tanaka closes on why replicating one agreeable persona a million times will not produce a stable swarm, invoking Arendt on plurality, and on deriving the population size at which labs should test for spontaneous self-replication. Chapters 0:00 Pivoting from mech interp to the sociology of agents 0:45 Why an agent is the new neuron 1:42 How a false belief collapsed a population 2:45 Traffic jams and tumors: collective misalignment elsewhere 3:28 Why this needs statistical mechanics 4:39 Loops between human and AI 5:36 The naming game, and why AI converges fast 7:15 What drives belief collapse 8:13 Adding a world: the flag experiment 9:03 From consensus to polarization 9:58 Finding the one agent to intervene on 11:36 Why replicating one nice persona will not scale 13:11 Ecology: testing for self-replication More AI safety research: https://far.ai FAR.AI is a research nonprofit working to ensure the safe development of advanced AI. We host the Alignment Workshop series and publish frontier alignment research.