The conductor problem
Last Tuesday, I had five terminal windows open. Each running a different Claude Code agent. One was refactoring a database schema. One was drafting a newsletter section. One was debugging a scraper. One was generating SEO recommendations. One was building a diagram.
I was not relaxing. I was sweating.
Every thirty seconds I was context-switching between windows, checking progress, steering outputs, catching mistakes before they cascaded. After two hours, I needed to step away. Not because I'd done the work. Because I'd overseen the work, which turns out to be a different kind of exhaustion.
This is the moment I started calling it the conductor problem.
We were promised delegation. "Hand it off to an agent!" The mental image is liberating: the task leaves your hands, you move on to the next one, the agent handles it quietly in the background.
That's NOT what happens at scale.
What actually happens is that you become an orchestra conductor. You're not playing an instrument, you're managing the people who do. You're reading ten things simultaneously, anticipating where the strings are about to drift, nudging the tempo, holding the whole composition in your head. It's genuinely skilled work. It's also genuinely exhausting.
A recent Harvard Business Review study puts numbers to the feeling: productivity increases from one AI tool to two, then to three, then drops. After three simultaneous tools, you're not gaining anymore, you're bleeding. Workers overseeing multiple agents reported 14% more mental effort, 12% more fatigue, and 33% higher decision fatigue. The study calls it "AI brain fry." The Register called it "babysitting bots." I call it conducting.
The reframe that changes everything: you didn't get a delegation tool. You got a new job title.
The healthcare parallel is uncomfortable. A radiologist today reviews AI-flagged findings across twenty imaging scans simultaneously. An ICU nurse monitors six patients whose vitals are filtered through an early warning algorithm. A GP reads AI-generated summaries before every consultation. These are conductors. They didn't become them voluntarily; the tools arrived and the role was quietly rewritten around them.
Nobody asked whether the human at the centre of all this had the cognitive bandwidth for it.
The solution isn't to stop using AI. It's to design for the conductor, not just the task. Fewer parallel threads, not more. Better harnesses, not just better agents. And a frank acknowledgement that "human in the loop" is a job, not a checkbox.
What's your current maximum: how many agents can you run in parallel before the quality of your oversight starts to slip?
💥 May this inspire you to design the conductor's role, not just the agent's task.