
1/3 I agree with Vitalik's point: if you restrict the set of coalition-level deviations the principal has to defend against, the principal’s problem gets easier. My question: can the restriction can be made strong enough to do the work we want from it without making the agents useless? Useful agents cannot be informationally isolated. They need models of the world, public signals, shared data, evaluators, tools, intermediaries, and other sources of context. But once agents inhabit overlapping information environments, restricting direct communication does not eliminate coordination capacity: common signals and shared context can themselves support tacit coordination. That creates a basic tension. The more we isolate agents to suppress tacit coordination, the less informed (and therefore less useful) they become. But once we give them enough shared context to be useful, we reopen informational pathways for coordination without explicit communication. The same information environment that enables useful action can also make tacitly coordinated deviations practically available. So perhaps the harder problem is not simply whether agents can communicate, but where the enforcement boundary falls within the broader information environment? That is not giving up on collusion resistance, but instead asking how to achieve it given informational interdependence.
It's possible that a killer app of adversarial governance mechanism design theory will end up being AI safety. Compare: vitalik.eth.limo/general/202… aiprospects.substack.com/p/p… There's a deep duality between the two environments. Both are about a less-sophisticated principal trying to get ideal outcomes from a set of more-sophisticated agents: in the first case, the principal is a static algorithm and the agents are humans, in the second case, the principal is humans plus weaker LLMs and the agents are stronger LLMs. A key finding was that you can achieve much better outcomes if you can guarantee limits to how much agents can collude (see: Nash equilibria being abundant, vs. cooperative-game-theory "cores" often being empty). That argument should naturally transfer to this new setting.
















