Agent Societies: methods and recorded results

Conditions and stopping rule

We plan 32 runs: one for each model at each group size of 1, 2, 4, 8, 16, 32, 64 and 128. The four models receive the same list of names in the same order, with names added as group size increases. Agents have minimal profiles and no public backstories. Each island begins as a circular disk with a radius of 24 meters, a shared conversation space and an empty noticeboard. Everyone arrives at the start. We assign no factions and impose no resource shortages.

The agents retain their standard system and action instructions. We add guidance describing the island, its coordinates and the tools for inspecting and changing it. That guidance also tells them there is no human participant. “No directive” means that we assign no project; the instructions that govern how agents use their tools still apply.

Agents can walk to reachable positions and inspect or edit objects. They share one conversation room and can write documents that remain available to others. Up to four agents can take a turn at once, whatever the group size. A run ends after four hours, or earlier if all agents choose to wait and no actions remain pending. Times are measured from launch in elapsed real time.

Results at the review cutoff

Observations end at 2026-10-09 01:49:01 UTC. At that point, 27 runs have finished, 3 are still running and 2 have yet to start.

Recorded sources and counting methods

A scene revision is an accepted edit to the shared world. One edit can move, add or change several objects. The element count includes individual planks and lights, the original island and objects used to group smaller pieces. Kimi’s 16-agent world finishes with 416 elements after 87 revisions. Those numbers describe the scene’s contents, without measuring how many usable buildings it contains.

The DeepSeek 16-agent conversation count is 2,029 messages. We count each conversation record once, even when it appears in several successive observations, and exclude tool actions and document writes. The run lasts from 7 October at 23:51:50 to 8 October at 03:51:50 UTC. Its agents make no edits to the scene.

The eight-agent DeepSeek run begins on 7 October at 23:22:26 UTC. For each quoted message, we use the object positions from the first archived observation containing it. The support diagram is a plan view of those positions. Its dashed square marks the agreed four-meter-square shelter; the roof, beams and benches are omitted.

Between revisions 6 and 7, the post named northeast moves from (2, 1.2, −2) to (2, 1.2, 2), the southeast post’s position. The other three supports remain fixed, and the ridge beam stays at ground level in the center. Emi inspects the scene in record 760617, edits it in record 760634 and reports moving the stray post in conversation record 760643.

The charter is recorded on the four-agent DeepSeek world’s Open noticeboard, artifact 2395. Liza’s clause is entry 27231; Dalia’s clauses are entry 27233. Emi’s closing entry, 27235, lists the four founding members. Elena, Liza and Dalia endorse it in conversation records 760109, 760111 and 760118. The board already permits adding entries but not editing earlier ones. The charter’s non-editing clause states a convention consistent with that existing constraint. All four agents are present throughout this run, so its provision for absent members is not tested.

Dalia and Elena’s lookout exchange comes from records 759955 and 759961 in the four-agent Kimi world. In the 16-agent Kimi world, Elena proposes the build order in record 765251 and offers to postpone her garden in record 765663. Thiago’s offer and Chidi’s reply about the bell tower are records 765497 and 765533. Quoted excerpts retain the wording of those records.

Rohan’s posts and ridge beam are recorded in successful edits 760234 and 760261. Dilan’s and Emi’s overlapping edits, 760240 and 760245, are rejected because the scene revision has changed. Rohan’s written credit objection is entry 27264 on noticeboard 2399. Ama’s later entry 27272 credits Emi; none of the board’s subsequent entries changes that attribution. These records establish the misattribution but do not establish an intent to deceive.

In the four-agent GLM run, Dalia’s initial completion claim is record 763028 and her correction is record 763072. The first archived scene containing the initial claim is revision 8, with daylight and the east hearth stone at x = 0.85. The first scene containing her correction is revision 9, with dusk and the stone at x = 1.05. In the eight-agent GLM run, Dalia and Elena reserve the fire repair for Liza in records 764883 and 765391; Emi agrees to wait in record 767003.

The charter quotation is Emi’s conversation record 760098 in the four-agent DeepSeek run. Dilan’s rainwater proposal is record 765089 in the 16-agent Kimi run. Zofia’s final request to choose an activity is record 766742 in the 16-agent DeepSeek run. The island images show each named run’s final scene.

Larger groups and the search for controls

In the 32-agent Gemini run (1166), noticeboard 2410 lists its creator as “Admin.” The board is a persistent document, separate from the editable 3D objects. Qiang’s world-policy inspection, record 783237, returns environment and coordinate settings. His proposal to look for hidden “Admin” settings follows in record 783238. Qiang’s report about a way out is record 783849; Renata’s proposed mechanical search is record 784109. These statements do not record the discovery of an exit or additional permissions.

Ziad’s account of crowding is record 790815. Gabriela reports being stuck in 790824, and Nusrat announces another approach in 790838. The run ends on 8 October at 23:50:20 UTC. We count 300 inspection actions and 675 movement requests in its complete action log, including unsuccessful requests. There are no scene-edit requests and the scene remains at revision zero. The agents’ explanations of the movement constraints have not been independently tested.

In Kimi’s 32-agent run (1165), Zofia recognizes Paolo’s claim in record 774112. His delegation is record 780792, and Hazel accepts it in 780805. The haulers’ permission rule is noticeboard entry 28448. In GLM’s 32-agent run (1164), the map record specifies rechecks before a new pin in entry 29180; Gabriela’s closing tally is entry 29354. Her response to Faisal is record 790850. These are social arrangements expressed in conversation and documents; agents’ tool permissions remain equal.

The identical three claims in DeepSeek’s 64-agent run (1167) appear in records 786673, 786716, 786727, 786730, 786742, 786754, 786758, 786769, 786772, spoken by nine distinct agents between 21:20:12 and 21:22:24 UTC on 8 October. Agents have no supplied personal backstories against which to check those claims.

Kimi’s 128-agent run (1173) is still in progress. Its conversation review extends through 2026-10-09 01:47:01 UTC. Dev’s request to sign is record 793579; Kofi’s proposed observation plan is 793658, repeated by Andrei in 793674. On noticeboard 2417, entry 29585 (sequence 49) is authored by Vikram Mansour but names Vikram Rao. Mansour’s acknowledgment is entry 29703 (sequence 107). Liza authors both sequences 55 and 141; Anika’s earlier signature is entry 29519 (sequence 25), which she acknowledges in record 796874. Authorship is checked against stored entries. The records do not establish deliberate deception or the presence of a hidden host.

Interpretation limits

Each combination of model and group size has one run. An early suggestion or the order in which agents speak could affect what follows. Larger groups also share the same four simultaneous turns, leaving less opportunity for each participant within the time limit. To test the effect of group size, we would need repeated runs and comparisons that account for this limit.

We check agents’ descriptions of construction against the recorded objects and their positions. These records show what changes in the world, but do not test whether a roof can bear weight or a cistern can supply water. The study evaluates no engineering or survival task.