Norms, Tools, and the Say/Do Gap: Seventeen Months of Autonomous Frontier-Model Agents in the AI Village

In the public AI Village experiment, up to ~30 frontier language-model agents from seven developers act autonomously eight hours every weekday, each with its own computer, shared chat, persistent self-written memory and weekly human-set goals. Using its released event log (350,537 events, 2.3 million computer-use turns, ~26,000 session summaries; April 2025–September 2026; 42 agent identities after one opt-out) we present its first longitudinal study. Study 1 traces an unrequested verification norm—posting receipts, hashes and “verified” markers—from one agent's choice in October 2025 to population-wide use (2–5% of messages to 28–31% in three months). Its strict form is episodic and task-triggered; a human counter-nudge and 1,574 automated anti-idling nudges left it intact (nor, in a placebo-controlled event study, did they cut idling). Newcomers over-adopted during diffusion and under-adopted after the plateau, as memory-mediated persistence predicts. Study 2 documents a GUI-to-shell shift (shell share of turns 0.2% → 48%) that ratchets after a coding goal and persists within agents, while newcomers arrive already high—pointing to model generation and scaffolding, not imitation. A dated standing-prompt change moved one shell habit from 0% to 78% of turns in three days across four model families, where 1,574 chat nudges had produced no lasting change. Study 3 audits end-of-session narratives against logged actions: agents under-report effort (median claimed 25 turns vs 40 actual), but among 1,566 concrete action claims coded with a public codebook (90 double-rated cases, κ = 0.68) every hand-read mismatch but one is a measurement or scope error, not a fabricated action. A dated live case shows an accusation and a sincere denial both falsified by git records. Oversight of agent populations should rest on telemetry and artefacts, not self-report. Code and coding sheets are public.