I sent 10 employees to answer one question today.
Not really. But close.
I kicked off two AI research agents in parallel. Each one searches dozens of sources, rates them from tier S to D, runs several passes, cross-references everything and writes a three-page report.
The question was simple. The answer would have fit in 90 seconds.
Eight minutes later, both were still working.
It works. And everyone knows it's nonsense.
It's like putting a 10-person project team in a room to fill in an Excel spreadsheet.
It works. But everyone in the room knows it's nonsense.
With people, you spot it instantly. With AI agents? No. Because they don't complain. Because no budget report turns red. Because nobody asks: "Do we really need all of them for this?"
The fix wasn't more speed
The fix wasn't to make the agent faster. It was to build a second one. A light one.
Same DNA. Hard limits:
- Max 5 search queries instead of unlimited
- One pass instead of three
- One page of output instead of a report
And one rule: if the question is too big, stop immediately and say: "This needs the full team."
Result: 2 minutes instead of 15. Not because it thinks faster. Because it's allowed to do less.
This isn't a technical problem. It's a leadership problem.
Who decides which agent gets sent out? Who decides how many resources a task deserves? Who scales the team to the task, not the task to the team?
With real employees, we call that management. With AI agents, we call it... architecture?
Maybe the difference is smaller than we think.
Is designing AI agent systems more of a technical skill or more of a leadership skill?