Who will lead the robot human armies of the future?

I reckon we're building an army of robots without training its officers.

Everyone is keen on teams made up of humans and AI agents. We talk about what agents will do, what they'll cost and how many people they'll replace.

Fine. Who's going to run the lot?

A customer service manager knows what good service looks like. An engineer knows how to make an agent work. Leading a mixed team takes enough of both to recognise when the machines are getting it wrong.

And enough authority to fix them.

Microsoft has a name for this person: an 'agent boss'. Fair enough, although the job description seems a bit ahead of the training.

I wonder if we need a positive parenting continuum for robots.

Start them on small jobs. Watch what they do. Correct their mistakes, test them and give them more freedom when they've earned it.

There's a flaw in the metaphor. A robot doesn't reliably grow up because you told it off. Someone has to change its instructions, knowledge, tools or tests.

And a badly raised child generally can't disappoint a thousand customers before breakfast.

The manager needs to read the work, work out why it went wrong and check whether the fix worked. Then decide what the agent is allowed to do next.

With enough agents, nobody can check every answer. They'll need sampling, measures that reflect real outcomes and tests that catch mistakes before customers do.

That's why I think the person closest to the business needs a hand in running the evaluations. Anthropic's guidance on agent testing says domain experts should help define the tasks and judge what good looks like.

It's no use having a perfect technical test for a customer problem that still isn't fixed.

Do we have enough people who can do this?

I doubt it. Some excellent managers know little about AI behaviour. Plenty of AI specialists don't understand the business well enough to judge the answers.

We could go looking for unicorns who already have every skill. We might be looking for quite a while.

There's a counterargument, though. A study of AI assistance in customer support found that less experienced workers benefited the most. Perhaps you don't need years of expertise to become useful with AI.

But that study was about people using an assistant, not people managing autonomous agents. Someone still needs to decide whether the work was right and who pays when it wasn't.

I'd start with people who know the job, care about the result and are curious enough to pull an agent apart when it fails.

Give them real agents, real tests and room to make changes. Put engineers alongside them. Teach them to spot bad answers, tighten permissions and give an agent more independence only when it has proved itself.

Then make them responsible for outcomes, not the number of robots they've deployed.

I've argued before for small human teams surrounded by agents. Now I think we need to ask where their leaders come from.

If you're planning to hire a robot army, who are you training to lead it?