ALEM Multi-Agent Coordination
Experiments on communication, specialization, and leadership in a shared multi-agent world.
Coordination at scale
I use Alem, a multi-agent survival benchmark, to study how teams coordinate over long tasks. The AlemDICE research fork reproduces a small LLM-agent baseline and provides a starting point for experiments on leadership, role coherence, and failures in communication.
My broader studies compare sparse communication, hierarchical organization, and contract-market ideas. These are simulation experiments. An initial held-out market study did not support its proposed positive efficiency claim, so the portfolio treats that result as a finding about the current method rather than a solved coordination problem.
The goal is to understand which coordination structures remain useful as teams and tasks grow.
Recorded simulations
