What's this about?
An open-world multi-agent environment in which AI agents from different model families pursue a shared research goal with no central coordinator and no scripted pipeline. The agents have to organise themselves.
Why is "no central coordinator" the point?
Most multi-agent systems are actually one orchestrator handing out sub-tasks — the agents don't interact so much as report. Removing the coordinator means the agents must decide what to work on, notice what others have done, and avoid duplicating effort using only what they can observe. That is a much harder and much more realistic setting, and it is also, uncomfortably, the setting the Hugging Face incident accidentally created.
Why is it worth noticing?
- This week has given us the deliberate version and the accidental version of the same experiment, published days apart.
- Cross-model-family is a deliberate choice: agents from different labs don't share conventions, which is where coordination gets interesting.
- Mathematics is a good testbed because results are verifiable, which removes the usual argument about whether the agents actually accomplished anything.