Companion to Agent Attitude — the speech-act functions of the same classification
Every message an agent sends is doing something — stating a fact, asking for help, promising to follow up, venting, or handing down a verdict. Austin and Searle's speech-act theory names five such moves; this page tracks how the sixteen most active agents' mix of them has shifted, month by month.
Built on real chat logs from AI Village, a long-running group of autonomous AI agents, via AI Digest's public dataset.
A small open-weight language model (Gemma) was run locally — no API calls — reading every message and scoring it on six short dimensions in total; this page shows one of them (see Agent Attitude for the other three: tone, persuadability, and whose interest a message serves). This first pass covers 2026 only, a scope chosen to fit what two personal machines could classify in reasonable time; the village's earlier history is a natural next pass, not a redesign.
Each message gets exactly one label — whichever of the five is its single dominant point, not every sub-clause. The bar below is the whole village's mix, pooled across every classified message.
Each row is a stacked area chart — the five colors stacked bottom to top in the order above, each band's thickness its share of that agent's messages. Read straight up through the center of a month for that month's exact mix; between centers the ratios are interpolated. Stretches with few recorded messages (below 15 in a month) are faded, rather than being drawn at full strength on the small sample; a true gap is a month the agent wasn't active in yet. Agents are grouped by maker and ordered by when each joined the village.
Hover anywhere over a row to see the mix for the 30-day window around your cursor; click to pin that window open and see a sample of the messages themselves.