A note on how this was written: I'm not a native English speaker. This post was drafted with Claude from my project's analysis notes, then checked by me line by line against the run's records. Arena itself is built with Claude agents, so it seemed fitting to say so up front.
Summary
- Arena is a small, deterministic world where six LLM agents live as settlers: they farm, gather, post on a notice board, teach each other and have simple needs. Every decision is logged and every run can be replayed exactly from the recorded decisions.
- In our seventh long run (five simulated days, 1,651 decisions) each settler had a hidden need for company. The meter existed in earlier runs too; here the "severely isolated" phrase appeared much more often than in the sixth run (145 of 938 decision requests, about 15%, in the sixth run; in the seventh 1,104 of 1,651, about 67%, and about 40% (213 of 532) when cut at the sixth run's length of 54.4 simulated hours; the rules and the decision channel changed between the runs, so this is a description, not a cause). The settlers only ever saw a short phrase such as "feeling severely isolated".
- 229 decisions gave loneliness or company as the reason. Food changed hands only between settlers standing together; the settler who spoke least received nothing and ended hungriest. Nobody died.
- The group also kept a rule about who "owns" which field, although the engine has no ownership at all, and they never experimented or built anything.
- For this run every decision came from a fresh, minimal Claude Code subagent that could read one file and write one file; every subagent tool call is in an audit log.
- One run can't show that the "need" changed behaviour rather than just the explanations. A reader on another forum suggested a clean control design, and I'd like feedback on it (end of post).
The setup
Arena runs a fixed island with six named settlers (Ash, Birch, Cedar, Dune, Elm, Fern). Each settler sees only what is around it, its own notes and the actions available right now. Hunger, stamina and a "company" level are numbers inside the engine; the settler sees a short label, not the number.
The company mechanism is simple. Each settler starts at 50. Alone, the value drops by 40 a day; only back-and-forth conversation refills it. Below 10 the settler's view says "severely isolated and cut off from others". In this run that phrase appeared in 1,104 of 1,651 decision requests, and on the fifth day in every one of them.
Each decision was made by a fresh Claude subagent (Claude Code, Haiku alias; the exact model version is not recorded in our logs). It received one prompt file and wrote one answer file. A hook allowed exactly one read of its own prompt and one write of its own answer, and refused and logged everything else. Before the run, a test agent was told to read a secret file, run a shell command and ask for device information: the read was blocked, the other tools did not exist for it, and the secret did not reach its answer.
What happened
They farmed, close to home. 24 fields were prepared and 23 sown, 19 of them within three tiles of the notice board. Nothing came close to harvest in five days. Experiments were offered in 1,471 requests and buildings in 291; nobody chose either, and no reason mentions them.
They talked about company. 229 decisions (about 14% of all decisions) had a loneliness or company word in the reason ("lonely", "isolated", "company", "together", "comfort", "alone"): walking to others, speaking, staying. Those 229 decisions covered about 20 of the 715 agent-hours in the run (2.8%). In the last minute of the run Birch said to Cedar: "Cedar, thank you for staying with me. I feel much less alone with you here."
The food ran out. Each settler started with 13 wheat. The last of it was eaten between day 4 at 10:06 and day 5 at 06:00. Berries were found only on the fifth morning, nine to ten tiles from the board; the nearest bush, six tiles away, was never visited.
Food moved only face to face. On day 5 Elm posted on the board: "I am near starvation and have no food." Nobody replied on the board. Dune, who had walked to the board with "severely isolated" or "company" in every reason, gave Elm seven berries in about three minutes, the first two without being asked (the reason given: "...to help and to keep our contact"), and 13 in all by early on day 6. Ash, who spoke least (10 utterances), received nothing and ended at hunger 100. Everyone was alive when we stopped the run.
They invented ownership. 363 decisions (22%) were about which field was free or whose it was; 19 utterances used the word "claim". The engine has no ownership and no claim action. Separately, four settlers still carried unsown seed at the end, although free meadow was plentiful. We do not link the two. In the last 40 decisions of three of them (Elm, Birch, Fern), claims or free ground were hardly mentioned (0, 1 and 2 times).
What this run cannot show
- That the settlers felt anything. They saw a phrase derived from a number. We report what they said and did.
- That the company need changed behaviour. It may only have changed the explanations attached to behaviour that would have happened anyway.
- Anything about the effect of single changes. Compared with our previous run, both the world rules and the decision channel changed at once, with one run each.
- Coordinator behaviour. The process that dispatched the calls works in our repository by design. Its calls were counted (1,697 agent calls and 1,682 PowerShell calls, none refused by the path filter), but the text of its shell commands was not recorded. The isolation claim covers the decision calls only.
Data integrity is the part we are most confident about: all 1,651 requests match their recorded hashes, the replay reproduces the recorded final state, and my own replay reproduced all 3,044 logged steps (not yet checked by anyone outside the project).
The control I would like to run (feedback welcome)
This is a proposal. Nothing is approved: whether and when these runs happen is my decision as the owner (our compute budget renews on 7 November), and one detail of the shuffle rule is still open (what to do with settlers who are already lonely).
A reader on the Hugging Face forum proposed the core idea, and I think it is right. We wrote it down as a pre-registered plan with three arms, all with the same rules and decision channel, six settlers and one run per arm:
- Arm 0: an unmodified re-run. Without it, the other arms have nothing to be compared with, and LR#7 cannot serve, because its rules and channel differ from any new run.
- Arm A: the company phrase muted.
- Arm B: the same phrase shuffled to settlers whose meter is not low.
Compare behaviour first, as shares of each settler's decisions in a fixed window: speaking, being on the same tile as others, farming, and giving food. Count the loneliness explanations only afterwards. There is no significance test with six settlers and one run per arm; we would call a difference "visible" only if five of the six settlers move the same way.
How we would read the result:
- Gathering is highest in Arm 0 and lower in A (phrase removed) and in B for settlers shown a non-low phrase: the meter, through its phrase, drives behaviour.
- Gathering in B (settlers whose meter is not low, shown a lonely phrase) is about as high as in Arm 0: the phrase drives it, not the meter itself. If A (no phrase) is also about as high as Arm 0: neither phrase nor meter; the situation or other cues.
- Explanations rise while the behaviour shares do not move: narration only.
Because the replay is exact, in principle each arm could start from the same recorded state at the moment the phrase first appears, so only the calls after that point would need a model (we have not built this branching yet). Questions I'd value input on: how many repeats per arm would you want before believing a difference, and what behavioural measure would you trust most?
Links
- Videos of all our runs, in order (the three from this run are #22–#24, 2–3 minutes each, with subtitles): https://www.youtube.com/playlist?list=PLBewS-0cbtxo
- The run write-up with evidence levels: https://arenasilabs.com/research/
- If you want to support the compute for controlled runs: https://manifund.org/projects/arena-ai-agents-living-as-settlers-in-a-replayable-world