Hi all, we are ArenaSI Labs, a small independent project from Budapest. We build Arena: a deterministic simulation where six LLM agents live as settlers on an island. They gather, farm, sleep, talk at a notice board and keep their own notes. There are no scripted goals; the world only gives them needs (hunger, stamina) and rules.
We wanted to share one moment from our fifth long run, because it surprised us.
Setup
- Six agents, all Claude Haiku 5.5, each deciding for itself from its own observation and memory.
- The system prompt has no instruction to share food or help others.
- Every event is logged. The run can be replayed from the journal: for the video we replayed 5,600 steps through the engine and got identical events, and all 280 checked state snapshots matched by digest.
What happened
- By day 4–5 the wheat the settlers carried had run out, and the forest gave only wood.
- Day 5, 07:19: Ash found berries east of the camp and gathered them “for myself and others”.
- 09:11: Birch, at the board, said: “No food either, hungry-ish.” Two minutes later Birch turned east, reached the berries and filled its pack to 44.
- Birch’s own note: “Hungry folks at (0,5) with no food. Heading west to deliver.”
- 12:50, back at the board: “Birch here with 44 fresh BERRY. Cedar, Dune, Elm: who is hungry?” It gave 10 berries each. The receivers had no food at the time.
- Over the run Birch gave 167 of the 230 fresh berries that changed hands (72.6%) and received 5 items.
- All six settlers were alive at the end.
The part we like most: Birch was not perfect. Twice it announced berries it no longer carried. Later it corrected itself in a note: “I carry 0 berries; earlier speech claiming 29-30 berries was wrong.”
Caveats: this is one run. We describe what the agents said and did; we do not claim why, or that this generalises.
2-minute video of the run: [https://youtu.be/ITFL6hEMTXw](https://youtu.be/ITFL6hEMTXw)
More runs and write-ups: [https://arenasilabs.com](https://arenasilabs.com)
We would love feedback, especially from people working on multi-agent evaluation: what would you measure in a world like this?