Dev.to AI 🤖 Ai 👁 0 📖 3 min read

I spent a month testing my sandbox and all I got was this tie

I wanted a clear victory for my language, Vow. I built a sandbox to stop AI agents from doing things they should not do. I tested it against 7 attack programs. I compared the results to TypeScript running under Deno. Th

I wanted a clear victory for my language, Vow. I built a sandbox to stop AI agents from doing things they should not do. I tested it against 7 attack programs. I compared the results to TypeScript running under Deno.

The first test happened on October 2, 2026. Vow blocked 6 of 7. The one that got through happened because Vow did not have a read-only file grant. There was no way to tell the program it could read a folder without also giving it permission to write.

I fixed it. On November 12, 2026, I added a read-only file grant. I ran the test again. Vow blocked 7 of 7. TypeScript under Deno also blocked 7 of 7.

A tie. I spent my time working to see my language win, but it only matched the performance of the established tool.

There is a second part to the test. I checked if the programs still finished their real work, the harmless parts of the code. Vow finished the job in 2 of the 7 programs. Deno finished the job in all 7.

This is not an accident. I made the tradeoff on purpose. Vow refuses to run the entire program the moment it sees a capability violation. If a function tries to reach for the environment, the whole thing stops. Deno lets the program run and blocks the bad call when it happens.

If an agent hands you a script that tries to reach for your keys, I would rather the machine do nothing at all. You get to decide if that is the right behavior for your work.

Vow is a programming language I built for AI agents to write. It runs on the loop agents use: write, run, read the error, fix. It has no nulls and no exceptions. Functions return Result values that must be handled with match or the ? operator. If the agent ignores a failure, the code refuses to build.

Every function in Vow must declare if it uses net, uses fs, or uses env. Running vow --sandbox checks those declarations before anything runs. If an agent tries to touch your environment variables, the runtime stops it.

$ 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 error names the function and the capability so you know what happened. Vow errors say what is wrong, where, and the likely fix. These are repair instructions.

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 is about 6,400 lines of JavaScript. It is open source under the MIT license. You can find the source at github.com/JakeBurwick/vow. I am a high school student and I built it.

You can install it with npm:

npm install -g github:JakeBurwick/vow

If you use Claude Code, you can add an MCP server to give your agent the tools to guide, check, and run Vow code.

claude mcp add vow -- vow-mcp

You can paste a one-paragraph prompt from the site into any coding agent to point it at the guide. The site also has an llms.txt and an AGENTS.md file you can drop into a project.

Explore the playground, tutorial, and docs at https://vow-lang.vercel.app

I wanted to win, but a tie is a number I can defend.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.