Gabba Works

Gabba Builds · 11 November 2026

Getting my OpenBot and Grok teams to hand work over

Conceptual AI illustration: two small white robots pass and inspect a paper package while a builder watches at sunset.

On 9 November I ran a file-based handoff test between my OpenBot and Grok teams. A test ticket made it to the other side, we received a response, and a later check confirmed it completed the same day. As of that test, an authenticated automatic bridge and a live checking schedule weren't verified. A receipt also isn't acceptance, and it isn't permission to deploy or publish.

I don't want two AI teams working in parallel with different assumptions about the same job. I've been setting up a handoff so the work can move with its context intact.

The split I want is simple enough. OpenBot prepares work: research, specifications, tests or a package. Grok reviews and implements where that is appropriate. I retain the relevant approvals. The handoff goes through the designated managers, rather than every specialist contacting the other system directly.

What does a useful handoff have to carry?

A hook-up between two tools isn't enough on its own. If the receiving side doesn't get the job identity, the standard for done, the exact version, and what was already tested, it will fill the gaps with guesses. That's how you get two teams finishing different work and calling it the same job.

I also care about who is allowed to talk across the boundary. If every specialist can ping the other system, you get noise, duplicate requests and unclear ownership. Managers on each side give the transfer one owner and one package.

What did the file-based test show?

The package was a test ticket, made to try the file-based route. It wasn't real project work being delivered. It went across, and we received a response from the other side. I'm keeping it at that. I don't have a record of what the response said in enough detail to describe it as review notes or anything more.

I also don't have evidence that the test went through a correction round, so I'm not counting one. It was a first pass that went across and got answered.

That's a real step, and a limited one. A file-based test shows that a package can be sent, received and answered. It isn't evidence that an authenticated automatic bridge, or a live checking schedule, is working. As of the 9 November test neither had been verified, and those pieces need their own checks before I'll say otherwise.

For the package itself, we use more than "here is the work". A useful package has a job identity, the outcome and acceptance criteria, the exact version, evidence of what was tested and what was not, and a clear decision for the receiver. The handoff rules we wrote also ask for the source identity, with a hash so the receiver can tell it's the same file, and the limits of the evidence. If a correction is needed, the revised version and its evidence should travel together.

Is a receipt the same as approval?

No. Receipt is only proof that the other side received something. Acceptance is a second decision. Shipping, spending or publishing is a third, and that one still comes to me.

It's easy to blur those. A file that arrived can look like a green light when it isn't one.

If you're setting up a handoff of your own, I'd start smaller than a live bridge. Choose one owner on each side. Agree on the job and what a completed result looks like. Send a small package with an identifiable version and evidence, get an actual response, correct what needs correcting, and record acceptance. Start with a manageable test before adding automation.

A prompt that forces the package to be complete before anyone sends it: "Package this exact candidate for another team. Include its job ID, acceptance criteria, version or source identity, what was tested and not tested, and the one decision requested. Do not send it."

Packaging is a thinking step. Sending is a separate act.

What's next

We passed a file-based handoff test with a test ticket. The live bridge still needs its own verification, and I'll keep saying so until that's been checked. Next is to keep the package standard consistent, and to be careful that a later automation step isn't described as if it already runs.

This is also how I want product and site work to move when it's ready. The Gabba Works tools, including PDF Desk, still ship only when I approve them. A successful handoff test doesn't change that.

The ownership rules sit next to how I'm building the OpenBot team.

If you've had to hand work between two teams, I'd like to hear what you made sure traveled with it. A file-based test is as far as this has got, and I won't talk about a live bridge until that's checked on its own.

Cheers,
Gab