The most incompetent workers I've ever hired

A few weeks in with AI sub-agents. The setup works. The judgment doesn't.

For the last few weeks I've been running an experiment. I own two product areas at work, international expansion and partner strategy, and I wanted to see what would happen if I gave each one its own AI sub-agent.

The idea wasn't original. There are dozens of posts on LinkedIn lately about people running "teams of 12 autonomous agents" doing real work in real companies. I never quite believed any of them, but I wanted to find out where the truth actually was.

So I asked a cloud agent to set me up with two workers. It called them shift workers, which I liked. Each one got a paragraph describing the work area, what it owned, and what success looked like. For partners, that meant building context on which companies we partner with, how they use our APIs, and what we're trying to enable. For international, it meant tracking the regions we've launched in, the ones we're considering, and the cross-functional dependencies for each.

Then I gave them access to the same set of inputs:

* My email, both incoming and outgoing, filtered to anything related to their topic

* The Slack channels where their work stream is being discussed, plus the general product channels for context

* Meeting notes and transcripts, both my own and the team summaries that come through email

* A shared signals database that both workers write to, so each can see what the other is seeing

* A short list of the key humans involved in each work stream, so they know who matters

Every few minutes, each worker checks for new signals on its topic. When something happens that looks like it might move the work forward, the worker drafts the next thing. Sometimes that's a planning doc. Sometimes it's a message back to a partner. Sometimes it's a note for me to take into a meeting.

The default is to move forward. The only times the worker stops and pings me are when there's an actual blocker, or when the next step is a communication going to someone outside the team. That second one matters a lot. I don't want my agents emailing partners or external stakeholders without me reviewing first.

And honestly? The setup works. The output is reasonable. The drafts are decent starting points. I'm getting some time back.

But.

Here's where I am, three weeks in: it feels like I've hired the most incompetent workers I've ever had.

They ping me for everything. My Slack inbox is a constant stream of approval requests, most of which don't really need approval. Many of them don't even make sense. And the underlying problem isn't volume, or speed, or context. It's that these workers are pathologically agreeable.

A real PM, even a junior one in their first year, knows things that have nothing to do with the immediate request in front of them. They know the company is trying to grow margin this half. They know we made a call three months ago to consolidate on a particular API pattern. They know our partnerships team has explicitly said no to one-off custom integrations. So when a partner sends a message saying "hey, we should build the API in this XYZ way," a junior PM reads it, checks it against everything they know, and pushes back without me having to step in.

The agent doesn't do that. It reads the suggestion, weighs it on its own merits, finds it perfectly reasonable, and starts drafting an enthusiastic response about how excited we are to explore this direction. Sometimes I catch it. Sometimes I almost don't.

The gap isn't capability. The agent can write the response. It isn't context either, since it has read every relevant Slack channel and email thread for weeks. The gap is judgment about when to disagree.

That gap, I think, is the actual hard problem with autonomous agents right now. Not "can they do the work." They can. The question is whether they understand the broader objectives well enough to know when an incoming request actually conflicts with them. Right now, in my experience, they don't. They are eager and willing and they will cheerfully nod the company in the wrong direction if you let them.

I don't know when this gets fixed. Six months, maybe. A year. Possibly two. I don't have a strong prediction. What I do know is that the version of "AI workforce" you read about on LinkedIn, the one where someone has 12 agents running their entire business while they go to the beach, is not the version I'm actually able to build today.

What I am able to build is a shadow team. Workers that draft at the speed I can review. They're useful. They're not autonomous. And they reach out to me way too much.

If you're trying to set up something similar, I'm happy to share more about how it's wired up. The architecture is simple enough that anyone with a Slack workspace and some willingness to wire up scripts can do it. The hard part isn't the setup. The hard part is everything that comes after, when you realize what these workers still can't do.

Send me a note, or comment on the LinkedIn version of this post. I want to know what failure modes other people are hitting.