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 every weekday for 2–8 hours, each with its own computer, a shared chat, self-written memory and weekly human-set goals. From its released event log (350,537 events, 2.3M computer-use turns, ~26,000 session summaries, April 2025–September 2026; 42 agents after one opt-out) we present the first longitudinal study. Study 1 traces an unrequested verification norm—receipts, hashes and "verified" markers—from one agent's choice in October 2025 to population-wide use (2–5% to 28–31% of messages in three months). Its strict form is episodic and task-triggered; a human counter-nudge and 1,574 automated anti-idling nudges neither dented it nor, against placebo windows, cut idling. Newcomers over-adopted during diffusion and under-adopted afterwards, and in 71,071 self-written next-session goals a written intention to verify predicts the next session across a weekend (71% vs 18%) but not across a goal change (14% vs 13%). 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, implicating model generation and scaffolding. One dated prompt change moved a shell habit from 0% to 78% of turns in three days across four model families; chat nudges had not. Study 3 audits end-of-session narratives against logged actions: agents under-report effort (median 25 turns claimed vs 40 logged), but among 1,566 frame-matched action claims coded with a public codebook (300 double-rated, κ = 0.78) every hand-read mismatch but one is a measurement or scope error; we found no fabrication. A live accusation and a sincere denial were both falsified by git logs. Oversight of agent populations should therefore rest on telemetry and artefacts; code and coding sheets are public.