An agent writing a language it was never trained on

George Djabarov · August 8, 2026

The goal I'm chasing: software built with no person in the critical loop, specified in human language instead of written as code. You specify what you want, precisely. A system of agents and tooling builds it. The spec is what you create and maintain; the source is just an intermediate step.

I've started calling it no-code AI programming — not the drag-and-drop kind: the programming still happens in full, a human just doesn't write or read the code.

As a proof of it, I built a programming language this way. Promise — compiler, standard library, toolchain — built by that system while I wrote no code at all: I set the goals, the spec and the design decisions, and filed bugs when things broke.

The language is optimized for an AI to write, not for a person: fit what an agent can hold in limited context, make errors less likely, and surface the rest as early as possible.

Which leads to an interesting question: can an AI write software in language it was never trained on?

So I gave an agent a task with nothing to recall from: 456 words of specification → ~650 lines of Promise. The task was mini-git — a content-addressed version-control tool, ten subcommands, state on disk. One prompt, no intervention, ~20 minutes.

• It did not compile on the first try. No training data to fall back on, so it learned the language from the toolchain — and the compiler corrected it, several times.
• Once it type-checked, it ran correctly on the first execution. Every subcommand. Binary files round-tripping byte-for-byte — which it tested thoroughly, on its own initiative. No logic bugs.

Many compile errors, zero logic errors. Types, ownership and exhaustive matching turn "the model guessed wrong" into a message on the terminal instead of a bug found in production three weeks later. The agent doesn't need to have memorized the language — it needs one that tells it when it's wrong.

The full run is public — watch it stumble its way to a working solution (prompt and source are linked from the recording).

Promise is early and not production-ready, and honestly I wouldn't recommend spending time learning it. But the toolchain installs from a single download if you want to play. You can read the code, deliberately, for when you need to audit or debug something. You just shouldn't have to.

What's next: I've left the language mostly unattended for six weeks now, and I want to let it stabilize. Then a few key modules. Then the real test — building Reactor, the open-source orchestrator for BASE, in Promise itself: the system that builds software, written in the language it built.

Until then, the Zoo is where the evidence goes. If you have a task you think would break this, send it to me — I'll run it and publish whatever happens, failures included.

First published on LinkedIn on 8 August 2026.