Languages for AI agents

Programming languages whose intended author is a model, not a person. Three camps, by what they optimise. Taxonomy follows agentlanguages.dev, the independent catalogue of the field; this is a selection, not the full index.

“Active” means at least one commit on the default branch in the 90 days before 2026-08-02, via the GitHub API. This field moves in bursts, so a quiet month is not the same as abandoned - but a project at 0/0 has not been touched in a quarter.

Syntactic

Cut the tokens a model spends writing and reading code: dense aliases, minimal punctuation, one canonical form.

Language	Design move	Active
ilo	Short fixed builtin aliases; design changes argued against measured token cost; agent-session failures filed as language bugs	active
Mog	Embedded-only, flat operators (no precedence), host-granted capabilities; full spec in ~2.7k tokens	quiet
Codong	One canonical form per operation; JSON errors carry fix and retry fields	quiet
X07	No text at all: programs are canonical JSON ASTs, edits are RFC 6902 patches	active
Magpie	SSA as the surface syntax; every value typed at definition; LLVM-backed	quiet
NERD	Every operator replaced with an English keyword; MCP client primitives	quiet
LLMLang	Prefix-arity AST, De Bruijn indices, compile-time linear ownership	active
Lume	Conventional surface plus token-budgeted docs retrieval (lume kb pack --max-tokens N)	active

Verification

Spend tokens to buy certainty: machine-checked contracts, effect types, solvers between generated code and production.

Language	Design move	Active
Zero	Agent-first systems language, tiny binaries, JSON diagnostics with repair plans	active
BAML	Close the escape hatches: runtime types with no any or unchecked casts, typed errors with exhaustive matching, schema-aligned parsing that repairs malformed model output instead of rejecting it. Compiler work began Nov 2025; the older prompt DSL was rebranded BAML v0 in Jul 2026, so the two release tracks share a repo	active
Hyperlambda	Programs are node trees rather than text, so generated code arrives as an executable AST that is validated before it runs and only whitelisted invocations execute. The containment is original rather than retrofitted: whitelist-invocation.hl dates to Dec 2022, while the agent positioning and the MCP server are 2026	active
Vera	Mandatory contracts verified by Z3; LLM inference as a typed effect	active
Thermite	req/ens/fx contracts discharged by Verus, Lean and Z3; five-level assurance ladder	active
AILANG	Row-polymorphic effect types, no loops; effects carve IO/FS/Net/Clock/AI	active
NanoLang	Shadow test blocks required per function; core proved with Coq theorems	active
Vow	Machine-checked vows discharged by ESBMC bounded model checking	active
Aver	Functions carry intent, effects and a verify block exporting to Lean 4 / Dafny	active
Intent	Mandatory pre/postconditions checked by Z3; one source targets Rust, JS, Wasm	active

Orchestration

The language coordinates agents rather than replacing their code: workflows, capability grants, deterministic replay.

Language	Design move	Active
Fabro	Workflow harness; .fabro files are Graphviz digraphs, a CSS-like sheet routes models	active
Boruna	Capability-safe deterministic workflows with hash-chained evidence bundles	active
Plasm	Path expressions over typed API graphs with dry-run plans	active
Lumen	Markdown-native sources with algebraic effects and compile-time determinism	quiet
Marsha	English functional files compiled to tested Python by an LLM (2023; the category’s ancestor)	quiet

Corrections welcome at hello@ilo-lang.ai. See also ways to be token efficient.