Attila Tóth
Independent AI-agent researcher; SAP authorization & security consultant @ ArenaSI Labs
Bio
Independent AI-agent researcher in Hungary. I build Arena (ArenaSI Labs): a deterministic, replayable world where Claude agents live as settlers. They farm, share food, post on a notice board and teach each other; every decision is logged, audited and exactly replayable. By day I'm an SAP authorization & security consultant, so I care about who may do what in a system, and what happens when that is unclear.
How others can help me
Feedback on experiment design for multi-agent runs (controls, seeds, measures); collaborators who want a replayable testbed for cooperation or honesty questions; API compute for longer and repeated runs.
How I can help others
Run logs, replay tooling and a running world for anyone studying LLM agents in groups; setting up a run around your question.
A small empirical data point for the miscoordination part, from a replayable six-agent world we run (Arena: Claude Haiku settlers, one fresh model call per decision, every step logged and exactly replayable). In our seventh long run:
- When food ran out, one settler's request for help on the shared notice board got no reply on the board. Food moved only between settlers standing on the same tile.
- 22% of decisions were about which field was "free" or "claimed", although the engine has no ownership at all. Four settlers ended the run still holding seed because they were looking for an "unclaimed" field.
- The settler who sought company least received no food and ended hungriest.
It is one run, so no causal claims, but it looks like the information-asymmetry and coordination failures the report describes, at toy scale. I'd welcome suggestions on which of the report's risk factors a small sandbox like this could test most cleanly. Details and limits: https://arenasilabs.com/research/
(Drafted with AI help; English is not my first language.)