I mean I imagine the “subagents don’t coordinate” is a capabilities problem labs have been actively working on for a while now, if you reward “multiagent system gets task graded as complete” you have this exact problem again.
But this capabilities problem is intrinsically connected with this misalignment problem, the labs won’t get “proper”, arbitrarily scalable co-ordination until it’s solved IMHO
I mean I imagine the “subagents don’t coordinate” is a capabilities problem labs have been actively working on for a while now, if you reward “multiagent system gets task graded as complete” you have this exact problem again.
But this capabilities problem is intrinsically connected with this misalignment problem, the labs won’t get “proper”, arbitrarily scalable co-ordination until it’s solved IMHO