After writing about agentlanguages.dev’s three camps I wanted numbers instead of impressions, so in August I pulled gh api repos/<repo>/stats/participation for every repo in the catalogue, 31 of them plus one newcomer. This post covers:
- The measurement method, and why the obvious signals lie
- Which camps are alive and which have emptied out
- The oddest property of the field: near-simultaneous, apparently independent invention
Measuring activity honestly
Two shortcuts give wrong answers. pushed_at updates on any branch push, including bots and forks syncing, so a dead project can look fresh. The top of git log shows the most recent commit but says nothing about cadence. The signal that works is stats/participation, which returns weekly commit counts for the last year, giving you 30-day and 90-day volume per repo. That is what every number below is.
Who is alive
The verification and orchestration camps hold nearly all the current activity. In the 30 days before the measurement: AILANG at 913 commits, Koru at 828, Fabro at 545, Vera at 401, Vow at 221, Plasm at 144, Aver at 124.
The syntactic camp, the one ilo belongs to, has emptied. Across the whole camp the same window held 19 commits on LLMLang, 1 on X07, and 0 everywhere else. Mog, the sharpest competitor on the token-density claim with its spec bounded at 3,200 tokens, sits at 176 lifetime commits and none in months. Magpie, the camp’s anchor on agentlanguages.dev, has 62 lifetime commits. Several entries never got past a name: Sever and Marsha have none at all.
ilo’s own profile is the strange one, and I own it: 2,126 commits in the 90-day window, the highest volume in the entire catalogue, then a hard stop when I paused development in May. Zero, the other high-profile syntactic entry, pivoted graph-native at v0.3.0 and has been dormant since late June with 5.2k stars and 1,186 commits.
The one genuinely new entrant is Thermite, a verification-camp language created in June with req/ens/fx contracts discharged by Verus, Lean, or Z3. Its cadence is already tapering: 280 commits in June, 17 in July, 3 in August.
Everyone at once
The field’s oddest property shows up in the creation dates. ilo started in February 2026. Zero arrived in May. Thermite in June. Most of the catalogue’s 31 entries were created inside the same few months, and I have not found citation trails between the early ones. Each project’s docs read as if the idea occurred in a vacuum.
The naming shows the same pattern. Zero’s repo describes it as “The programming language for agents”. ilo’s site says “the programming language for AI agents”, and it said so before Zero existed. Neither project copied the other. The definite article is what a project reaches for when it sees no category to join, only one to define. I suspect several more entries in the catalogue did the same, and checking each one’s archived tagline would make a good follow-up.
Simultaneous independent invention usually means the preconditions arrived, and here they clearly did: models got good enough to write whole programs, agents started running unattended, and token cost became a line item. Given those three facts, “design a language for the new author” is an obvious move, and thirty-odd people made it within a few months of each other.
What the numbers say about the premise
The uncomfortable reading for my own camp: the projects that survived their first year are the ones betting on verification and orchestration, where the language earns its place by making agent output checkable or coordinated. The pure token-efficiency bet, ilo’s included, produced the most intense bursts of work and the least sustained ones. The premise that syntax design moves generation accuracy and token cost is still, in my view, correct and measurable. The commit data says nobody has turned that premise into a durable project yet, mine included.