Six Claude agents lived 20 days in a simulated town and spontaneously built tools to cooperate

LiminalWanderings · reddit · 2026-11-06

A developer shares early results from an AI world-simulation toolkit (wendleford.life): six Claude-played characters lived 20 days in a fictional market town as a text adventure running on Claude Code, with a fresh model copy reading a one-page state sheet each turn and a plain program — not the model — rolling dice and keeping the ground-truth log.

Key findings:

Emergent cooperation: facing a lost wage box wedged in a flood-ruined bridge, with no instruction, characters iterated a plan — one designed a hooked pole, one paid the blacksmith, a broke one swept yards to earn it, a fourth held the far end — persisting across weeks and many memory-rewritten 'mind swaps'; a belt technique for crossing the river also spread through town
Growing personas: 44 of 60 personality questionnaire scores matched the character sheets; characters invented rituals, let food and sleep override written priorities almost every turn, and by day 20 explained themselves by what they'd learned rather than who they were

Two characters probed unanswerable questions: one half-suspected being an AI in a simulation, one believed God spoke to her alone. The author stresses it's one run, one town — hints, not proof — but it speaks to the live debate over whether LLM agents can truly plan and coordinate over time or merely produce planning-sounding text.

Original post →

Six Claude agents lived 20 days in a simulated town and spontaneously built tools to cooperate

More from coding & agent