AIL

August 2, 2026 · View on GitHub

English · 한국어 · 中文 · for AI agents

A programming language being built from scratch for LLMs — deliberately excluding human reading, writing, and learning.

Why a new language?

Every enforcement mechanism we have lives on one of two sides of a gap:

  • Outside the language: agent harnesses, permission systems, sandboxes, CI gates. They block actions without understanding what the program intends.
  • Inside the language: type systems, Rust's ownership, capabilities, effect systems. They enforce rigorously — but against memory and type errors, not against the risks of the agent era.

The intersection — in-language enforcement against agent-era risks (irreversible actions, technical debt, LLM misuse, runaway loops) — is empty. AIL is built to stand exactly there.

This follows from HEAAL (Harness Engineering As A Language, pronounced "heal"): don't recommend against what must not happen — make it inexpressible or unrunnable through the language's grammar, parser, and runtime. The language itself is the harness. Full text: docs/HEAAL.md (Korean).

What is different?

An AIL program is not a description of procedure. It is a contract of intent, with three mandatory parts:

  1. What it is trying to achieve — declared purpose
  2. What observation counts as success — decidable success criteria
  3. What is forbidden — the boundary of capabilities, effects, and resource budgets it owns

A program missing any of the three does not parse. And what it discards is exactly what existed only for human cognition: readability syntax, learning curves, ergonomics, comment culture. What it gains: token efficiency, structural uniformity, machine-checkable claims, context self-sufficiency. Details: docs/AIL.md (Korean).

Current status

Concept stage — honestly. The philosophy (HEAAL) and the language concept are codified. Grammar, parser, and runtime are not designed yet; any mention of them anywhere in this repo is direction, not specification.

How this repository works

This repo runs on gil (GIt for Language model): every piece of work is recorded as chain (purpose) > cycle (problem) > steps (define → hypothesis → verify → analyze → conclusion) in the commit graph. Not just conclusions — hypotheses, falsification conditions, and failed branches are all preserved.

Note the symmetry: a gil cycle demands purpose, success criteria, and falsification conditions — the same three-part contract AIL demands of programs. This repository already runs on the principle of the language it is building.

Contributing & documents