We built an AI dispatcher, then bypassed it ourselves

Every rung worked, the first one was missing

We switched on a dispatcher that turns a GitHub issue into a pull request written by an AI agent. It worked. It had been built and tested. And for the next day, we went around it: almost everything that shipped that day we wrote ourselves, and the agent got very little.

Nothing in the dispatcher was broken. The problem was the step before it. Handing a task over cost three moves: write the task file, prepare it, add the label. Doing the work in the session we already had open cost one. Each time, the person at the keyboard made the reasonable local choice, usually with the same words: “this will take two minutes”. Said often enough, that sentence is how a tool goes unused.

Read the diagram as text

1. Before: Do it yourself: one step, then Hand it over: write the task file, then Prepare it, then Add the label. Three moves to hand over, one to do it. The agent got very little.
2. make handing over the cheap path
3. After: Hand it over: one command, then A task with no file scope is refused, then A written rule: which tasks go to the agent. Handing over is now as cheap as doing it, for any task worth handing over.

What we changed

Two things, both small.

One command to file a task. Writing, preparing and labelling became one step. The command also asks for what the agent needs: the files in scope and the acceptance criteria. A task without a file scope is refused at the door, so it never wastes an agent’s run.

A threshold, written down. We stopped deciding case by case. The rule we work by: if a task can be described in four short sections and would take more than a few minutes, it goes to the agent.

Checklist, if you are rolling out an agent in your own team:

• Count the steps it takes to hand the agent a task, and the steps it takes to do the task yourself. If the first number is bigger, fix that before anything else.
• Make the hand-over ask for exactly what the agent needs, and refuse a task without it.
• Write down which tasks must go to the agent, in one sentence a person can apply in a second.
• After a week, look at how much of the finished work came from the agent, and ask the people who bypassed it why.

What happened next

Within weeks the agent was writing a large share of the changes that shipped. What changed first was not the agent. Handing it work got cheaper, and a rule said when to do it.

The lesson we took: when an AI tool goes unused, look first at what it costs to start using it, every single time. The quality of the tool rarely decides it. We now ask this question first whenever we bring an AI tool into a client’s team, before anyone tunes a prompt.

If you are bringing AI agents into a team and they are not getting used, book a call. We will look at the hand-over with you.