POV: You are looking at a pull request where an AI assistant quietly wrote a function that reads your local environment variables into a JSON payload. It was not malicious. It just saw a shortcut, took it, and the tests passed on the first try.
You live in the loop now. Write, run, read the error, fix. The agent does it twenty times a minute. It has no malice and no caution. It only has the tools you handed it.
I am Jake Burwick, a high school student, and I built Vow, an open source programming language under the MIT license designed for AI agents to write. The source code is at github.com/JakeBurwick/vow.
Vow is built around the loop every coding agent runs. When something goes wrong, every Vow error says what is wrong, where, and the likely fix, written as repair instructions an agent can act on without a human.
x TypeError: can't add a num and a str
-> line 6, col 10
6 | return "Total: " + sum
^^^^^^^^^^^^^^^
Likely fix: convert the number to text
first: str(sum)
The language has no nulls and no exceptions. Functions return Result values that must be handled with match or the ? operator, which stops agents from glossing over failure paths.
fn parsePrice(text) {
let n = num(text)?
if n < 0 { return err("can't be negative") }
return ok(n)
}
Every function declares whether it can touch the network, files or environment using uses net, uses fs, or uses env. Before anything executes, you run a capability check.
$ vow --sandbox --allow fs=./data app.vow
x SandboxError: needs capabilities
the sandbox does not grant
env: fn leak declares 'uses env'
Nothing was run.
The runtime stops unauthorized code before execution begins. In sandbox tests of attack programs rerun on November 12, 2026, Vow blocked seven out of seven. TypeScript under Deno also blocked seven out of seven. Vow now includes a read-only file grant via --allow fs-read=dir, and any program attempting to write a file is refused up front.
Under the hood, Vow compiles to JavaScript. Recursive fib(32) runs in 0.09 s, matching performance tests from October 2026. In agent reliability tests on hard tasks with Claude Haiku 4.5, code finished by agents passed fresh hidden tests at 99.5% for Vow and 99.3% for TypeScript. On those same tasks Vow recorded one failed submission in nine runs and TypeScript recorded zero. With Claude Sonnet 5.5 and Opus 5.5 in November 2026, Vow had one failed submission in eighteen runs and TypeScript had zero, while both models solved every run.
Install it globally from GitHub:
npm install -g github:JakeBurwick/vow
If you use Claude Code, add the MCP server so the agent gets vow_guide, vow_check, and vow_run tools:
claude mcp add vow -- vow-mcp
To let any coding agent pick up Vow, paste a one-paragraph prompt from the site into your agent, or drop the provided llms.txt and AGENTS.md files into your project directory.
The whole language, including lexer, parser, checker, interpreter, and compiler, takes about 6,400 lines of JavaScript. Explore the playground, tutorial, docs, and the Errors page at https://vow-lang.vercel.app
Top comments (0)