Taking the Agent on Board

On Coffee with Digital Trailblazers, we spent a good part of the show sorting AI agents into boxes: augmentation, human in the middle, agentic. Useful boxes. But somewhere in the middle of it, it occurred to me that we were debating the wrong thing. We already know how to do this. We do it every time we hire a bright young person.

You do not hand a new hire the keys on day one. You give them a small job. You check the work. If the work holds up, you give them a bigger job. If it drifts, you correct it fast. And once you trust them, you stop hovering, but you never stop auditing. That is onboarding. It works for people. It works for agents. The only real difference is that the agent never forgets a step and never gets tired, which is exactly why the checking still matters.

I have run the full range of this myself, so it's not theory.

I started with my old blog posts. I turned an agent loose on them: it picks one, writes a teaser, and posts it three times a week. At the outset, I was checking each one. The format came out well, so I loosened the rope. I still look in once or twice a week to make sure the posts are within the guidelines and still sound like me. You audit it like any other process, and you correct it when it drifts.

After becoming comfortable with that, I explored the other end of the spectrum: board prep. Of course, provisions are in place to maintain the confidentiality of that material. I have an agent that reads board material, reconciles it against minutes from prior meetings, and builds the context around what has been happening. It makes observations. It makes recommendations. But I still make every decision.

It will always be my prep, but it will never get my board seat.

So here is my recommended progression: 

1. Start small. Give the agent one job you can check.
2. Check the work. Not the theory, the actual output.
3. Widen the rope when the work earns it. Trust is built in stages, not granted in a policy.
4. Audit on a schedule. Once or twice a week beats hovering every day, and it beats discovering a problem a month late.

The question is not which box the agent belongs in. The question is whether you have trained it the way you would train anyone you trust with your name on the work. Agents can remember more and build context faster than a young employee. In many ways, they are better at executing routine tasks. But they lack judgement and are not better at knowing what matters. That is still your responsibility.

Train it like an employee. Trust it in stages. Check it on a schedule.

Captain Joe 

Follow me on X @JPuglisiLLC