Labs / Systems · Updated 2026-09-11

How Lehvel's AI systems hand work to each other

We got tired of carrying prompts and status between AI systems by hand. So we built a shared system that does it, with GitHub as the record.

By Founder

Published

Running experimentBuilt and running, being proven.
  1. Submitted
  2. Picked up
  3. Working
  4. Checkpoint saved
  5. Checked
  6. Complete

Explore with AI

What we asked and what we found

Every entry answers the same eight questions, in this order.

Question

Can two different AI systems hand work to each other without a person carrying every message?

Why it matters

One AI system plans, another does the work, and a person reconciles the two. When the person is the messenger, the company moves only as fast as that person can copy and paste, and every dropped message is a failure nobody sees.

State

Running experiment Built and running, being proven.

What we built or tested

  • A record of every task and its status, with a separate queue for each AI system, so a task is delivered once and a failed one is retried or set aside.
  • GitHub as the record of what was asked and what was done. Chat is for watching the work, never for deciding it.
  • A rule that the AI never chooses which client site or project it changes. A trusted source sets that.
  • Live checks that tell five states apart: received, picked up, in progress, saved and confirmed complete.
  • A forced-failure test confirming each task ran once, work restarted cleanly after a crash and a blocked task stayed blocked.

Evidence

How work flows
A request arrives, the coordinating system assigns it, a worker does it, the result is checked, and the outcome is recorded back to GitHub.
Crash test
Work was raced, retried and restarted across separate processes against shared state.
Choosing models
The cost-per-finished-task experiment ran on this system.

What we learned

  • The record has to live where a person can read it and an AI can act on it. A project repository does both. A chat thread does neither.
  • Picked up and confirmed complete are different states, and most failures hide between them.

Limitations

  • Reliability claims cover only the checks we ran. This is not a fully autonomous company.
  • This is Lehvel's own system. Clients receive the work it delivers, not a login.
  • The first version was paused on September 17, 2026.

What changes next

Set cost and reliability limits for each kind of work, then publish the status vocabulary as the evidence format clients see.
Failed tests are published as results. A state on this page describes today, not plans.

Next in Labs

  • Next: Running operations by voice

    Owners do not update forms. We are testing whether an owner can say what changed and have it become a change to the business record.

    Experiments

← All of Labs

Explore with AI

Get Upstream

One email a week: what we asked the AI, what it got wrong, and how we checked.

Our research reaches 17,000+ followers across our channels. Upstream is the weekly version.

Weekly. Unsubscribe anytime. Privacy