agentlanguages.dev is a community catalogue of programming languages designed for AI agents to write, maintained by Alasdair Allan, who also builds the Vera language. It grew out of his essay “Three camps alike in dignity”, and the camp structure it imposes is the most useful classification I have seen for this space. The shape:
- Syntactic: strip ambiguity from the syntax itself, so generation goes right more often
- Verification: make code mechanically checkable, with contracts, proofs, and SMT solvers
- Orchestration: constrain how agents coordinate, runtime over language
- An adjacent section for work that does not classify cleanly
The bar is intent
The listing criterion is explicit: the designer must target LLMs and agents as the authors of the code. Tools that merely call models at runtime are out, and so are languages whose only agent story is Copilot compatibility. That is the right bar. “An LLM can write it” is true of every language with enough training data. “Designed to be written by one” is a commitment you can check against the spec.
Anchors ground each camp: Magpie for syntactic, Vera for verification, Boruna for orchestration. Submission is a pull request adding a markdown file to src/content/languages/, with evidence and a self-classified camp.
Where ilo sits
ilo is a syntactic-camp language by this taxonomy: token-minimal text, positional arguments, prefix notation, a spec written for the model rather than the human. It is not in the catalogue. That is my omission rather than the catalogue’s, and the fix is the documented one, a PR with the evidence attached. The manifesto and the token benchmarks are exactly the kind of “designed for agents as authors” evidence the bar asks for.
Reading the syntactic camp as a group is sobering. The premise all of us share, that generation accuracy and token cost respond to syntax design, still gets argued repo by repo rather than measured across the camp. A shared catalogue at least puts the claims on one page where they can be compared.
What the map leaves out
One boundary the taxonomy does not draw: languages designed for agents to write versus formats designed for agents to read. Skills files, AGENTS.md conventions, and spec-shaped docs are the reading side, and they are converging on markdown while the writing side fragments into dozens of syntaxes. The catalogue scopes itself to the writing side, which is fair, but the reading side is where adoption has been faster.