3DVR experiment · Human v0

Programming should feel more human.

Human Language Lab explores a human-first, agent-native programming language: express intent clearly, keep the meaning strict, and compile it into ordinary inspectable software.

Open the experiment Read the first program

The idea

Instead of spending more compute teaching agents to navigate huge, ambiguous codebases, make the programming language itself smaller, clearer, safer, and easier for both people and agents to understand.

app GarageCRM

data Customer:
    name
    email
    status = "lead"

when Customer is created:
    send welcome email
    notify sales

when payment succeeds:
    Customer.status = "client"
    create project

Design rules

Human-readableCode should communicate intent before syntax trivia.

Machine-verifiableAmbiguity should become a compiler error, not a guess.

Agent-nativeSmall structured programs should reduce searching, rewriting, and retries.

InspectableAI can help author programs, but AI is not required to run them.

Current experiment

v0 is working.

The prototype includes a tiny .human syntax, a dependency-free compiler to Python, a Garage CRM example, tests, CI, and a source-size comparison. The first toy program is roughly 44% smaller than its comparable hand-written Python version. That is only an early source-surface measurement—not yet a claim about energy or total agent cost.

Next question

Can coding agents implement the same change in Human with fewer tokens, retries, mistakes, and wall-clock time than Python or TypeScript? That benchmark is the next step.