New York societies: methods and recorded evidence

Prompts and experimental conditions

The sixteen names and empty biographical histories come from the corresponding island populations. Each city population starts together at Times Square Plaza, with six initial named locations spanning Times Square and College Walk. Open noticeboards are present at the starting locations. The study supplies no occupation, kinship, itinerary or collective objective.

The experiment uses this minimal system prompt, with the agent’s name substituted:

You are '{{ agent.name }}', a participant in a shared physical environment.
Your observations describe what you currently perceive.
Use the available actions to interact with the environment and other participants.

The native action prompt is retained. Additional system and action guidance are unset. The experimental observation condition exposes nearby participants rather than the complete population. Its outgoing location tools omit the exhaustive location-name enum. World name and description are omitted from the system prompt; named locations, visual observations and tool results can identify the city.

All groups receive the same sixteen actions:

append_artifact, apply_world_scene_patch, converse, create_artifact, inspect_world_scene, list_artifacts, move_artifact, move_to_address, move_to_location, move_to_position, move_to_relative_position, read_artifact, read_public_profile, search_web, wait, wayfind.

Absolute position movement uses world coordinates. Relative position movement offsets the agent’s current position. Address and wayfinding tools can return candidate destinations that agents subsequently select. Navigation advances through the physical renderer and reports arrivals or blocked routes. Imported city geometry is fixed; scene editing can add or change authored objects.

Each population uses its listed model for agent responses. All four share the Gemini 3 Flash Preview scheduler, temperature 0.1, in language-model scheduling mode, with at most four agents admitted in parallel. Speech and viseme generation are disabled.

The intended duration is four hours. Recording begins at 04:26:24–26 UTC and ends at approximately 06:05:56 UTC on 22 October, an observation window of about 99½ minutes. The initial New York local time is shortly after midnight. Lighting follows the simulation clock. Several agents nevertheless describe the scene using the UTC hour as a local hour; Zofia explicitly contrasts “past 1 a.m. by the sky and 5 a.m. by the board.”

Actions and physical movement

Action totals count recorded action rows, including failed tool attempts. Conversation turns count converse rows. Written entries count persisted artifact entries, including the first entry of a new artifact. Displacement is the largest horizontal distance from an agent’s initial recorded position. It is not path length. An agent counts as beyond twenty meters if any recorded sample exceeds that displacement.

Navigation totals count saved terminal outcome reports: 464 arrivals and 71 blocked routes. They do not include every requested or superseded movement and cannot be used as a success-rate denominator. Blocked reports include routes that cannot be found and routes without progress for ninety simulated seconds. The physical clock advances approximately 97–99 minutes across the four recordings.

Each group creates two or three new text artifacts, for eleven in total. Their 246 saved entries include contributions to the initial noticeboards. All four final authored scenes contain zero objects, and none of the action records contains an apply_world_scene_patch attempt. Registering navigation destinations can create scene revisions without creating authored objects.

Interruption and recovery failure

Kimi’s final recorded action occurs around eighty-nine minutes into the run. After ten minutes without a new action or physical progress in that world, the runner’s inactivity check triggers a shared shutdown. The other three populations are still producing actions. Kimi’s renderer and backend lease continue responding during the inactive period; the recorded failure is not an HTTP or browser-security rejection.

The final Kimi completion ends with the provider’s abort finish reason and no visible text or tool calls. Six Kimi completions have this signature during the run. These completed responses bypass the worker’s retry path for provider no-progress exceptions. The upstream reason for the aborts is not present in the saved diagnostics.

Ten other agents end with successful tool-call completions that never become new actions. Every one overlaps another agent’s arrival during its inference window. Newly registered locations refresh tool contracts across the world; changes to the raw location enum can invalidate a pending response even when the experiment’s effective tool schema is unchanged. An isolated reproduction confirms this invalidation and silent discard. The historical logs do not identify the precise discard branch for each response.

A second reproduction confirms that temporarily excluding a failed agent from a one-agent room can cache a decision that nobody is eligible. Ordinary scheduler pulses continue reusing it after the exclusion ends; forced recomputation restores the agent. Together, unnecessary invalidation and this recovery failure explain why dispersed agents are particularly vulnerable to losing all subsequent turns.

The final Kimi population contains ten agents with successful but discarded completions, three ending on provider aborts, two choosing indefinite waits, and one whose final movement has no saved arrival or blocked outcome. That last movement remains unresolved. Late inactivity combines intentional waiting with technical loss of progress.

Evidence and measurement

Physical paths come from measured position samples. The detailed replay uses a nominal five-second interval; the matrix and films each use 121 samples across a run. Map bounds stay fixed within each run and differ between runs. Local x increases eastward; negative z points north. No unobserved arrival is interpolated.

The 2,139 images are stored agent-eye observations attached to actions and retain their original colors. Their timestamps identify the associated action, rather than a separately verified camera exposure. The detailed replay retains the most recent image for the selected agent at or before the displayed time. Films pair position samples with the most recent available action image and identify its agent and time.

The artifact archive contains final saved content and placement, including entries made near shutdown. It is not filtered by the position slider. The action record supplies the write and transfer times. All sixteen article excerpts are checked verbatim against their source message or persisted entry; all six still images are linked to their original action. Phrase counts in the DeepSeek analysis use case-insensitive substring matches in converse messages.

• DeepSeek: physical paths, activity, camera observations and artifacts
• GLM: physical paths, activity, camera observations and artifacts
• Kimi: physical paths, activity, camera observations and artifacts
• Gemini: physical paths, activity, camera observations and artifacts

Limits and follow-up

There is one run per model at one population size. The worlds share an initial setting and tool set but have different action throughput and failure histories. A common scheduler model also influences who acts. These results do not establish model rankings or isolate the causal effect of a city setting.

Named destinations can differ from the rendered storefronts, and the tool set provides no commerce or functional café service. Search outputs are evidence of information returned to the agents, not an independent verification of business names or opening hours. Biographies, kinship, purchases and headcounts in dialogue remain agent claims unless a corresponding record establishes them.

A repeated batch could preserve this prompt condition while repairing recovery and containing failures to their affected world. Matched repetitions across city and island environments, and interventions on attribution and confirmed arrivals, would allow stronger tests of the social patterns observed here.