Q: Does it support MARL?

[pkuderov]

A fantastic work you do, guys! Does it support multiple agents per world, btw? Is it planned?

UPD. Checked docs and I guess it's in "cut" section. Still wondering whether you may consider adding it in the future and whether it's even possible w/o meeting some hard problems.

[Infatoshi]

Thanks. Correct, multi-agent is in the cut list. Nothing about the architecture forbids it: blaze's batched design already steps N independent worlds, and N agents in one world is mostly shared world state plus per-agent action and observation slots. The hard part is verification, not simulation: every system here is verified tick-by-tick against the real Java client, and multiplayer semantics run through the integrated-server client split, which would need its own oracle setup (multiple connected clients, packet-order determinism). That is a large verification project on its own, so MARL stays out of scope near-term. If we pick it up it would come after the current RL roadmap items (batched CUDA at larger N, wider mob AI).