Agentic linguistics studies language through, with, and on autonomous agents. Its founding experiment suite — population convergence, the evolution of grammar, and a live in-context learning probe — is embedded in this page, deterministic to the bit, and will hand you the same numbers it handed us.
Consensus in agent societies obeys a power law. Across 30 seeded runs (N = 10 → 400), convergence time scaled with exponent 1.40 (R² = 0.994) — approaching the published N3/2 law for the naming game.
A random 16-word lexicon, transmitted through a learning bottleneck under communicative pressure, reorganized itself into shape-prefix + color-suffix morphology. Structure ρ = 0.98 ± 0.04, expressivity 99%.
Learnability pressure alone built structure (ρ = 0.70) but collapsed 40% of meanings into ambiguity. Only learnability × expressivity yields language-like compositionality — replicating the two-pressures account in a fully deterministic model.
Every simulation below was executed once, in a sandboxed reference environment, and frozen. The identical algorithms — same pseudo-random generator, same operation order — ship inside this page. Press reproduce anywhere and your browser recomputes the entire suite, then checks its results against the reference checksum. Agreement is bit-for-bit or it is failure.
N agents, no vocabulary, no coordination. Two meet at random; the speaker utters a word from its inventory (inventing one if empty); on failure the hearer memorizes it, on success both collapse their inventories to the winning word. That is the entire model — the minimal naming game of Baronchelli et al. (2006).
From these three rules, global consensus self-organizes: the lexicon balloons, peaks, and then collapses onto a single shared word in a sharp transition. We measured time-to-consensus for N = 10…400, five seeds each, and fit the scaling law.
Result — consensus time scales as t ∝ N1.40 (R² = 0.994), within reach of the published N3/2 law. A statistical law of language, recovered from 60 lines of code.
1000·s + N for s = 1…5. Speaker and hearer drawn uniformly without replacement per interaction. Convergence = all agents hold exactly one identical word (checked every 50 interactions). Exponent from OLS on ln t̄ ~ ln N. The hero console shows the N=100, seed-42 trace: distinct words in circulation (red) and rolling success rate over the last 200 interactions (blue), consensus at t = 4,550.Sixteen meanings (4 shapes × 4 colors), a lexicon of random two-syllable words, and a chain of learners — each acquiring the language from just 8 of 16 utterances of its teacher, then teaching the next. A minimal model in the Kirby tradition, with three learner types:
Rote memorizes what it hears and guesses the rest. Learnability-only generalizes: it aligns syllable positions with meaning features. Both-pressures generalizes and refuses ambiguity — no two shapes may claim the same prefix, no two colors the same suffix.
Result — rote never finds structure (ρ ≈ 0.00). Learnability alone finds it (ρ = 0.70) but degenerates: 40% of meanings become homophones. Both pressures together produce a fully compositional, fully expressive language — ρ = 0.98, expressivity 99% — within five generations.
| CONDITION | STRUCTURE ρ (GEN 30) | EXPRESSIVITY | VERDICT |
|---|---|---|---|
| Rote control | +0.002 ± 0.063 | 93% | expressive, structureless — unlearnable |
| Learnability only | +0.699 ± 0.165 | 60% | structured, degenerate — ambiguous |
| Learnability × expressivity | +0.983 ± 0.037 | 99% | compositional — language-like |
7001–7010. Structure = topographic similarity: Pearson correlation between meaning-space and signal-space Hamming distances over all 120 pairs. Expressivity = distinct signals ÷ 16. The ambiguity-avoidance rule in the third condition stands in for communicative pressure, following Kirby, Tamariz, Cornish & Smith (2015).Sixteen meanings, eight syllables, zero design. Rows (shapes) evolved shared prefixes; columns (colors) evolved shared suffixes — a morphology no one wrote. This language is the specimen probed live in Study C below.
The reference run serialized every convergence time and every structure curve into a canonical string and hashed it: FNV-1a = 0x4be15eab. Your browser re-executes all sixty simulations with the same seeds — typically in under a second — rebuilds the same string, and re-hashes it. One flipped bit anywhere, in any of the ~10⁶ simulated interactions, and the hashes diverge.
This is the founding methodological commitment of agentic linguistics: a result you cannot rerun is a rumor. Every experiment this programme publishes ships as data + seed + executable method, embedded where the claim is made.
Benchmark contamination is the quiet crisis of AI evaluation: if a test language exists anywhere online, a model may have memorized it. Study C makes contamination structurally impossible: the target language is the one that evolved in Study B, born from seed 7001 inside this page. It exists in no corpus, no archive, no training set — anywhere.
Press run and a live frontier model (Claude) receives K attested word–meaning pairs and must produce the words for four held-out meanings — exactly the in-context task facing agents documenting a real low-resource language from a minimum viable dataset. Scoring is automatic, against the evolved ground truth, at whole-word and morpheme level.
Deployment note: inside Claude, this page can call the model directly. When hosted at agenticlinguistics.com, route the same request through a one-file proxy holding your API key — the page degrades gracefully and everything else (Studies A & B, verification) runs fully offline.
Astronomy had the telescope. Biology had the microscope, then a model organism. Linguistics has had neither: no device that holds an entire language in view, and no organism on which the forbidden experiments — sealing off a speech community, rerunning a contact event, watching grammar change under controlled pressure — could ever ethically run.
That changed. There now exist artificial systems that use language rather than merely count it — and they compose into agents that converse, remember, form populations, execute a field linguist's workflow, and learn a language they have never seen from a sketch grammar placed in front of them. Agentic linguistics is the field that treats this as science, and holds itself to the standard demonstrated above: claims that arrive with their own reproduction.
An agent that transcribes, segments, glosses, and queries around the clock collapses the labor bottleneck that has defined descriptive linguistics since Boas. The linguist's judgment stays; the drudgery goes.
Populations of conversing agents are to linguistics what Drosophila is to genetics: not the real thing — a system where hypotheses about drift, contact, and complexity can finally be intervened on, rerun, and falsified. EXP-001 is the proof of principle.
Agents learn new languages in context — from a wordlist, a sketch grammar, hours of recordings. The unit of survival becomes the minimum viable dataset, not the billion-token corpus. Study C is that thesis, measured live.
Each pillar takes a branch of linguistics and asks what becomes possible when autonomous agents enter it — as subjects, instruments, or teachers.
Populations of memory-bearing agents conversing across thousands of generations — a laboratory for language change. Seal a community and measure drift; stage contact and watch for creole-like simplification; vary population size and test the linguistic-niche hypothesis under controlled conditions.
An orchestrated agent team that runs the documentation workflow — transcription, segmentation, interlinear glossing, lexicography, grammar sketching — with a native speaker verifying every claim. No pretraining on the language. No fine-tuning. Everything learned in context from a minimum viable dataset.
The dataset that documents a language can teach it. An agent that learned a language in context becomes a patient, always-on conversation partner — for heritage learners of a language with twelve remaining speakers, or any classroom whose language has no textbook and no market.
Agentic linguistics names a convergence three decades in the making — and answers it with a research programme and a reproducibility contract.
Steels' Talking Heads: embodied agents negotiate shared vocabularies from nothing. Language emergence becomes experimental.
Kirby, Cornish & Smith show structure emerging as language passes through human learners — the bottleneck made visible.
Park et al.'s generative agents form believable social worlds. LLM populations become plausible model organisms.
MTOB: an LLM translates Kalamang — unseen in training — from a single grammar book in context. The minimum-viable-dataset era opens.
Agent-driven corpus analysis arrives; the threads converge. This page names the field — and runs its founding experiments.
A field is founded by results, not declarations. Phase one is already live on this page.
Agentic Linguistics began with an observation: the tools that will decide whether thousands of languages survive this century are being built by people who have never glossed a sentence — and the people who gloss sentences rarely get to steer the tools. This programme sits deliberately in the gap.
Its standard is on display above: no claim without a seed, no result without a rerun button, no benchmark a model could have memorized. If you work on agents, on documentation, on revitalization — or you speak a language the internet forgot — the loop has a place for you.