Mo vs the agent-native cluster
One line: thirteen 2025–26 languages built for agents as authors: eleven from the agentlanguages.dev catalogue, plus Neam and NTNT, which aren’t in it. On the list to see which of Mo’s decisions they share (direction 1, direction 3, Q9, Q10) and what 42 separate attempts converged on. Format note: one page for many languages, so each gets a short entry rather than an idea list.
What it is (status as of Sep 2026)
The catalogue tracks 42 languages whose designers explicitly target LLMs or agents as authors. Tools that merely use an LLM at runtime are out of scope.[111] By camp: 17 verification, 14 syntactic, 7 orchestration, 1 adjacent, 3 unclassified.[111] It is maintained by Alasdair Allan, who also wrote one of the entries, Vera.[111] Most entries were first seen between February and June 2026.[111] Neam and NTNT both exist: public repos created in January 2026, with 7 and 8 stars.[114][112]
The ideas, one by one
Verification camp
- Intent (Go; v0.2.0, 5 stars). Every function has
requires/ensures, and loops carryinvariant/decreases. Anintentblock links a natural-language goal to contract clauses throughverified_by, and an unresolved link is a compile error. Z3 proves what it can; the rest becomes runtime checks.[111] - Vera (Python; about 2,400 commits, 9,382 tests, 413 stars live).[137] Mandatory
requires/ensures/effects, discharged by Z3 in the decidable fragment and as a runtime guard otherwise. Parameters have no names, only typed slots (@Int.0). LLM inference is a typed effect.[111]- Overlap / verdict: tiers and effects, yes. Reject unnamed slots: they cost human readability (direction 2).
- Thermite (Rust + Lean 4; June 2026). Mandatory
req/ens/fx, plus loopinv/dec. The Forge tool records per obligation whether a clause was proved, bounded-checked, runtime-enforced or trusted. “The project-level headline is the minimum”, and counterexamples stay failures.[111]fn sum(xs: &[u32]) -> u64 req xs.len() <= 1_000_000 ens result == spec_sum(xs) fx pure- Overlap / verdict: the closest to Mo’s contracts plus capabilities. Steal per-obligation evidence, with the minimum as the headline, for
verified:(Q6).
- Overlap / verdict: the closest to Mo’s contracts plus capabilities. Steal per-obligation evidence, with the minimum as the headline, for
- Vow (Rust, self-hosting; v0.2.0).
vowblocks and loop invariants go to the ESBMC bounded model checker before codegen. JSON diagnostics blame the Caller (requires) or the Callee (ensures,invariant). The compiler writes a Claude Code skill “generated from the same compiler version”.[111]- Overlap / verdict: Q9 plus crash reports (direction 21). Steal the blame field and the version-matched skill.
- Aver (Rust; v0.21). Every function has a prose intent (
?), declared effects (!), and a colocatedverifyblock. That block runs as samples, as hostile checks, or exports to Lean 4 or Dafny, and “the four readings can disagree”.[111]- Overlap / verdict: same-file tests (pick 12). Already have. Aver shows the risk of several verification readings of one block.
- AILANG (Go; 110 releases, 26 stars at cataloguing, 34 live).[137] Pure functional, with row-polymorphic effects. Capability categories (
IO,FS,Net,Clock,AI) are granted at launch with--caps. No loops. The compiler itself is written by AI agents.[111]- Overlap / verdict: capabilities (direction 15); Bosque’s no-loops move (bosque). Open: operator-granted capabilities at launch.
- Hale (Rust; May 2026). One primitive, the “locus”, replaces class, module, actor and service, and loci talk over a typed topic bus. Contracts range from
@no_syscallup to program-wideforbid reaches(A, B), answered with countermodel witnesses. No async colouring, lifetimes or locks. An MCP server ships in the compiler.[111]- Overlap / verdict: processes (direction 14) and
flows(...)(pick 9). Open:forbid reachesas the general form offlows.
- Overlap / verdict: processes (direction 14) and
Syntactic camp (and the verification crossovers)
- Tacit (Rust; v0.7.7, 3 stars). The AST is authoritative and serializes to exactly one canonical text. Definitions are BLAKE3-addressed, names live in a JSON sidecar, and malformed code becomes typed
Holenodes instead of a parse failure.[111]- Overlap / verdict: the inverse of Q10, which keeps names in text and IDs in a sidecar. Reject the sidecar names. Open: typed holes, so diagnostics keep working on partial files.
- Zero (Vercel Labs; C bootstrap; v0.1.1, 3.3k stars at cataloguing, 5,360 live, last commit 30 May 2026).[137] Stable diagnostic codes (
NAM003), typed repair plans (zero fix --plan --json), and guidance served by the CLI and pinned to the installed version. Capability objects onmain, no hidden allocator. A “pre-1 experiment”.[111] - Codong (Go; v0.1.3, 67 stars at cataloguing, 73 live, no activity for five months).[137] One canonical function per task. Nine bundled modules and zero dependencies. JSON errors carry
fixandretryfields. Compiles through Go.[111]- Overlap / verdict: a batteries-included stdlib (Q11), Q9. Already have.
- Axis (Rust; 3 stars, one initial commit). Dormant: last push 28 May 2026, four days after launch, and no benchmark has ever tested its small-model claim.[137] Aimed at 1B–7B models: twelve constructs for a backend, an LL(1) grammar, and grammar-aware logit masks shipped as JSON. “Failure is syntax”: missing auth and unindexed queries are compile errors.[111]
- Overlap / verdict: laws (direction 4). Open: constrained decoding, as on moonbit.
Outside the catalogue
- NTNT (Rust; created Jan 2026). “Intent-Driven Development”:
.intentscenario files checked byntnt intent check, plusrequirescontracts. 480 built-in functions, including auth, databases and background jobs. String interpolation is borrowed from Ruby. It claims “running production apps” but warns of breaking changes.[112][113]- Overlap / verdict: humans state intent and agents implement (direction 19). Already have. Its stdlib reaches well past Q11’s ceiling.
- Neam (C++; created Jan 2026). A compiled DSL for building AI agent systems: LLM providers, RAG, multi-agent orchestration, governance, running on a bytecode VM.[114]
- Overlap / verdict: none with Mo’s core. Claude’s reading: it uses LLMs at runtime, which the catalogue excludes.[111] Reject as a design peer.
What 42 attempts converged on
- Checkable beats correct. Verification is the largest camp.[111] The catalogue itself notes that nine entries enforce something per function, so “the live differentiator sits in what discharges the obligations rather than in their being compulsory”.[111]
- The same three layers as Mo’s Q8: mandatory contracts, a solver or model checker for what’s decidable, and runtime checks for the rest (Intent, Vera, Thermite, Vow).[111]
- Diagnostics built for agents: stable codes, JSON, and repair plans (Zero, Codong, Vow, Tacit).[111]
- Toolchains that ship their own agent guidance: skills or MCP pinned to the compiler version (Zero, Vow, Hale).[111]
- Explicit effects or capabilities in most verification-camp entries (AILANG, Aver, Thermite, Vera, Zero, Hale).[111]
- Rare: content addressing (Tacit) and no loops (AILANG). A rough keyword pass over all 42 entries (loose: a mention isn’t a feature) found content hashing in 4 and constrained decoding in 2, against contract language in 28.[111]
What it gives up
- Readability for humans. Vera drops parameter names, and Tacit drops comments and free formatting.[111]
- Stability and safety. Zero says breaking changes and security vulnerabilities “should be expected”.[111]
- Maturity. Most are single-author projects with single- or double-digit stars at cataloguing. Axis arrived as one commit.[111]
Evidence
- Size of the field: 42 catalogued, by camp as above.[111]
- Engineering signals: Vera’s 9,382 tests, Codong’s 1,427, AILANG’s 110 releases.[111]
- Adoption, live stars on 12 Sep 2026 (Robert’s agent-languages run): Zero 5,360, Vera 413, Codong 73, Aver 60, Thermite 53, Hale 35, AILANG 34, Vow 8, Intent 7, Tacit 4 (Phase 6 frozen), Axis 3. No usage data found.[137][112][114]
- LLM benchmarks (corrected: the first pass found none). VeraBench runs 9 models on 60 problems. Vera averages 98.7%, “ahead of Python’s 96.7% and a point behind TypeScript’s 99.7%”. The README itself notes TypeScript has “no headroom”, so there “the benchmark is measuring the models rather than the languages”.[149] On a 5-model zero-training subset, AILANG scored 96.8% and Aver 92.4%, against 97.0% for Python.[137] Conflict of interest: VeraBench is written and graded by Alasdair Allan, who authors Vera and maintains the catalogue. No third-party reproduction exists for any entry.[137][149]
- ilo, a self-reported failure. In ilo’s August 2026 CI baseline on Haiku 4.5, 1 of 13 scripted personas produced working code, 1 partial, and 11 failed. Its spec grew from about 16K to about 51K tokens. See case-against-new-languages.[137]
- Source caveat: the catalogue’s maintainer also authors Vera and VeraBench.[111][149]
What Mo should take from this
- Positioning: contracts, a solver and agent-facing diagnostics are now the camp’s consensus, not Mo’s edge.[111] Mo’s differentiation has to be elsewhere, and the v0 design doc should say so:
- spec altitude a human can actually read (direction 2, direction 26)
- processes with supervision and deterministic replay (direction 14, direction 21)
- capabilities as package permissions (direction 30)
- laws
- Proposal: a per-obligation
verified:record whose headline is the weakest obligation (Thermite).[111] - Proposal: contract-violation diagnostics carry blame,
requires→ caller andensures/invariant→ callee (Vow). This feeds the crash report in direction 21.[111] - Proposal:
moserves version-pinned agent guidance itself (Zero, Vow), alongside the MCP server proposed on unison.[111] - Question for Robert: should operators also grant capabilities at launch (AILANG’s
--caps), on top of narrowing inside the code? It is an input for Q17.[111] - Question for Robert: adopt Hale’s
forbid reaches(A, B)as the general form offlows(...)?[111] - Watch: Zero, the only entry with corporate backing and real attention, though its last commit was on 30 May 2026.[111][137]
Related
- language-landscape
- q08-verification-tiers
- q09-compiler-diagnostics
- q10-semantic-ids-and-editing
- q17-package-management-and-supply-chain
- moonbit
- unison
- bosque
Sources
[111] https://agentlanguages.dev/llms-full.txt — agentlanguages.dev full catalogue (llms-full.txt) [112] https://github.com/ntntlang/ntnt — NTNT: agent-native language with Intent-Driven Development (GitHub) [113] https://ntnt-lang.org/about — About NTNT [114] https://github.com/neam-lang/Neam — Neam: compiled DSL for AI agent systems (GitHub) [137] raw/research-runs/llm-authored-programming-languages.pplx.md — Robert’s research run: LLM-Authored Programming Languages, evidence base 2023–2026 (Perplexity, landscape prompt 3) [149] https://github.com/aallan/vera-bench — VeraBench: benchmark for Vera vs Python/TypeScript (Alasdair Allan)