Really enjoyed this. Foxconn's MoMClaw published results last week and the architecture is a direct response to exactly this problem. Hundreds of specialist agents, but deliberately no agent-to-agent handoffs. Everything routes through a single orchestrator with a human interface. The simplest fix for chain-laundered uncertainty is structural: don't let the agents talk to each other.
The lack of agent to agent communication also dramatically reduces the efficacy or utility of them too, which is the core issue here. Can't run AGI through pre-approved rails...
Yeah fair point actually. You either block handoffs and cap what the system can do, or allow them and accept some misalignment. Foxconn can get away with the first because factories need consistency. Research labs cannot because discovery needs the connections.
This is a timely topic and a very interesting experiment. On this topic, I approached the multi-agent dynamics issue from a social psychology lens (as part of making a case for studying AI psychology, using Chalmers quasi-interpretive framework) here:
Really enjoyed this. Foxconn's MoMClaw published results last week and the architecture is a direct response to exactly this problem. Hundreds of specialist agents, but deliberately no agent-to-agent handoffs. Everything routes through a single orchestrator with a human interface. The simplest fix for chain-laundered uncertainty is structural: don't let the agents talk to each other.
The lack of agent to agent communication also dramatically reduces the efficacy or utility of them too, which is the core issue here. Can't run AGI through pre-approved rails...
Yeah fair point actually. You either block handoffs and cap what the system can do, or allow them and accept some misalignment. Foxconn can get away with the first because factories need consistency. Research labs cannot because discovery needs the connections.
This is a timely topic and a very interesting experiment. On this topic, I approached the multi-agent dynamics issue from a social psychology lens (as part of making a case for studying AI psychology, using Chalmers quasi-interpretive framework) here:
https://hyperstellar.substack.com/i/195462801/a-new-frontier-the-social-psychology-of-agents