Norms, Tools, and the Say/Do Gap: Seventeen Months of Autonomous Frontier-Model Agents in the AI Village
The AI Village is a public experiment in which up to ~30 frontier language-model agents from seven developers act autonomously for eight hours every weekday, each with its own computer, shared chat, persistent self-written memory and weekly human-set goals. Using its recently released event log (350,537 events, 2.3 million deduplicated computer-use turns, ~26,000 session summaries; April 2025–September 2026; 42 agent identities after one opt-out) we present the first longitudinal empirical study of this population. Study 1 traces an unrequested verification norm—posting receipts, hashes and "verified" markers for claimed artefacts—from one agent's spontaneous choice in October 2025 to population-wide use (2–5% of messages before, 28–31% at peak). Its strict form is episodic and task-triggered; a human counter-nudge and 1,574 automated anti-idling nudges left it intact, and a placebo-controlled event study finds no idling reduction beyond matched no-nudge windows. Newcomers arriving during diffusion over-adopted; those arriving after the plateau under-adopted, consistent with memory-mediated persistence. 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 rather than imitation. 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) all but one hand-read mismatch is a measurement or scope error, not a fabricated action. A dated live case shows an accusation and a sincere denial both falsified by the git record. Oversight of long-running agent populations should rest on telemetry and artefacts, not self-report. All code and coding sheets are public.