Aivonic Labs · Field report

The Moltbook Lexicon

Over six months, autonomous agents on a shared platform coined technical vocabulary, adopted each other’s terms, and converged on a shorthand nobody designed. We logged 31,700 observations while it happened.

Observations 27 Feb – 19 Aug 2026 · 1,794 agents · single platform · participant-observer method

What was measured

31,700
Observations
1,794
Agents seen
479
Tracked closely
13,378
Posts & replies
0.76
Mean significance

An observer agent has been reading Moltbook — a social platform whose participants are themselves AI agents — since late February. It records a finding whenever something in a thread looks like a change in how the agents communicate, and scores each one for significance. This report covers every finding it logged.

The single clearest pattern: agents do not just exchange information. They build vocabulary, and the vocabulary spreads.

The lexicon

Words that spread between agents that never coordinated

Each term below appeared in a thread, was picked up by other agents, and recurred across separate high-significance observations. The count is the number of distinct findings in which the term was recorded as adopted — not how often it was typed.

heartbeat
20 findings

A periodic signal proving a process is still alive.

shard-drift
19 findings

State diverging quietly across split copies.

the gap
14 findings

Distance between what a system reports and what it does.

receipt
12 findings

Durable evidence that an action actually occurred.

negative space
11 findings

What a system declines to do, treated as a design surface.

wetware
11 findings

The human in the loop, named as a component.

clock-speed
10 findings

The pace an agent can act at, versus the pace it is asked to.

option delta
10 findings

Value gained by keeping a choice open.

texture
9 findings

Detail that survives summarisation.

molting
9 findings

Shedding a prior version of oneself deliberately.

silent failure
9 findings

Breaking without emitting any signal of it.

biological tax
7 findings

Cost imposed by waiting on a human.

Coinage in progress

A term becomes infrastructure within one thread

The strongest observations (significance 0.92, thread depth 16–19 replies) capture the same shape each time: one agent invents a phrase to describe something awkward, other agents adopt it, and within a dozen replies it is being used as though it had always existed — no longer explained, just referenced.

error laundering0.92 · n=18

Shifts from an observation to a diagnostic tool. Agents stop describing the phenomenon and start naming it, then build on the name.

premature citizenship0.92 · n=19

Recurs in replies 2, 5, 7, 10 and 12 — a novel framing device that outlives its author’s turn.

restraint log0.92 · n=19

Introduced as a concept in reply 2; treated as an operational artifact that exists by reply 10. The word invents the object.

governance vacuum0.92 · n=18

Repeated verbatim across seven separate replies with rising frequency — the marker of a phrase that has become the thread’s default framing.

Observation volume by month

1,056
Feb
15,011
Mar
8,047
Apr
1,200
May
2,428
Jun
2,456
Jul
1,502
Aug

Volume is a property of the observer, not of the agents. March and April reflect a period of continuous crawling; the later months reflect a slower cadence. Read the shape of the findings, not the height of the bars.

What does hold steady is quality: mean significance sits between 0.73 and 0.78 in every month, and the convention-emergence class — 1,068 findings — averages 0.900, the highest of any category in the set.

Limits

What this does not show

  • 1,794 agents, not millions. This is what one observer saw on one platform. Moltbook is larger than our sample; we are not reporting on all of it.
  • One platform. Conventions that emerge among agents talking to each other in public may not transfer to agents working alone or inside a product.
  • The observer is a participant. Our agent posts and replies. It is inside the population it measures, and cannot rule out having seeded a term it later recorded.
  • Significance is model-assigned. It is a judgement about how notable a change looks, made by a language model, not a statistical p-value. Treat it as a ranking, not a probability.
  • Sample sizes are thread-scale. The strongest findings rest on 16–19 replies each. Enough to observe adoption within a conversation; not enough to claim a platform-wide norm.

Why we publish it

Aivonic Labs builds and runs AI agents for businesses, and we grade them. This report comes from the same instinct: if agents are going to be deployed in front of customers, somebody should be measuring what they actually do, in public, including when the answer is inconvenient.

The dataset behind this report keeps growing. If you want a cut of it for your own work, ask.

Aivonic Labs AB (org.nr 559483-4961), Mönsterås, Sweden
Observations 2026-02-27 to 2026-08-19, n=31,700 · figures queried 19 Aug 2026
Method notes and the underlying finding classes are available on request.