Gabba Works

Gabba Builds · 11 November 2026

Building an AI team without making a mess of the work

Conceptual AI illustration: a builder and four small white robots assemble a wooden model house in a sunlit workshop.

I'm building an OpenBot team with roles for research, marketing, design, operations, independent review, skill engineering and management. Roles are useful, but they don't do the work of keeping jobs clear. The hard part is giving each task an owner, an outcome, limits, and a way to report what actually happened. We haven't measured a gain from the setup.

Having a role on the list doesn't mean each one is off running on its own at the same time as the others. I'm also deliberately not hanging this story on a team size. The useful question is whether the work is clear, not how long the roster looks.

Where does a team of agents actually go wrong?

Once you have more than a couple of agents, the problem usually isn't that nobody is working. Two agents can start the same job without noticing, or a result comes back with no owner, no evidence, and no sense of what was allowed. Naming roles doesn't fix that by itself.

I've been trying to make each task clear enough that the person, or the agent, receiving it knows what they're responsible for. That means an owner, an outcome, completion criteria, authority limits, and a way to report what actually happened.

Without those, a confident status is too easy to send and too hard to use.

What rules are we using so far?

The working rule is that a specialist can help without taking over the whole job. A reviewer can point out a problem without approving a release. A manager can coordinate without restarting research that has already been done. Some decisions stay with me: publishing, spending, deployment, and changes to account or security access.

These are working practices, not magic security controls. An instruction can tell an agent what it should do. The underlying access settings still matter. Telling a reviewer not to publish isn't the same thing as actually withholding the ability to publish.

Have we measured a productivity gain?

No. We haven't measured a gain, so I'm treating this as an experiment rather than a success story. That doesn't mean there isn't one. It means I can't show it, so I won't imply it.

A practical handoff, for a single task, looks like this. Give one person ownership. Describe the output and the check that would make it acceptable. Include only the context needed, state what the recipient can and cannot do, and ask for actual evidence rather than a confident status. Bring in a specialist when their knowledge helps, and retain the decision with the appropriate owner.

A prompt I've used: "Choose the owner for this task. State the expected output, acceptance criteria and authority. Reuse existing work, involve a specialist where needed, and report the evidence for the result. Ask only for material missing decisions."

Agents, like people, can reopen work that's already done if you let the brief drift. A good handoff reuses what exists and asks only for what is still missing.

I don't need a bigger team for its own sake. I need a team whose work I can understand, review and put to use. If I can't tell who owned a result, what "done" meant, and what they were allowed to do, the team isn't helping me yet.

What's next

I'll keep refining the way tasks are framed: owner, outcome, limits, and evidence. I won't describe the setup as proven productivity, and I won't pretend that written instructions enforce isolation on their own.

The file-based handoff between the OpenBot and Grok teams is one place those rules have to hold. This is the team I use while building the Gabba Works apps, including PDF Desk and the rest of the work on gabbaworks.com. Clear ownership on the team side is what lets me review that work before it goes out. It isn't a substitute for those reviews.

I'm still treating this as an experiment, with no productivity number attached. For now I'm just trying to keep the work clear enough that I can actually review it.

Till next time,
Gabba