I Had Seven AI Agents Running. My Attention Was a Mess.

I started one AI task. Then I waited. Then I opened a second tab. Then a third. Pretty soon I had seven or eight agents running at the same time.

Output went up. My attention did the opposite. I was bouncing between approvals, edits, and little decisions, and it felt productive because stuff was happening. It was the opposite of the single-tasking habits I’ve taught for years.

Once AI feels powerful, every task starts looking like something you should throw at it. You have a new hammer. Everything looks like a nail.

AI is not automatically productivity-positive. That’s the warning.

Some tasks are still faster, clearer, or more personal when you just do them yourself. Booking a familiar flight route. Choosing groceries when your preferences change in the moment. Those are not the jobs you should automate first.

Before you automate the next thing on your list, run a quick audit: AI or manual? If the task is familiar, personal, or changes in the moment, do it yourself.

Too many agents, too little thought

Brooks put words to something I was already feeling. When several AI-generated tasks are asking for approvals, edits, or decisions at once, you slip into reactive mode. The work still moves. Each decision just gets less thought than it would if you were actually focused on one workflow.

Does this sound familiar? You kick something off, get bored waiting, spin up another task, and now you’ve got a pile of half-finished drafts all wanting your attention.

Throughput can go up. Decision quality can go down. Running multiple agents may raise how much gets produced, but every handoff and review decision can suffer.

Set a personal agent concurrency limit for focused AI work. Start. Wait. Review. Decide. Then start the next one. Don’t constantly spin up more tasks just because you can.

The 80/10/10 rule

AI changes where focus is needed. It does not remove the need for focus.

I use an 80/10/10 rule for AI delegation.

The first 10% of the task is defining done, sharing context, clarifying resources, and setting the vision. The middle 80% is the AI doing the execution. The final 10% is human review, QA, and feedback.

That’s how you’d delegate to a teammate. Brief well. Let them work. Review carefully.

Skip the first 10% and you get vague output. Skip the last 10% and you get AI slop.

Even when an AI-assisted post, report, or email is technically accurate, if it smells generic, too perfect, or unlike the person who supposedly wrote it, people subconsciously trust you less. That’s a credibility tax. The skipped step is usually that final human 10%.

Use AI. Protect your trust.

Editing is still the job

For executive status updates, I let Claude create the first draft. Then I rewrite it so it sounds human, audience-appropriate, and non-technical. A technically thorough AI draft can still be wrong if it’s written for the wrong reader.

Editing is faster than authoring. You still have to edit.

Before you send anything, do a voice and audience pass. Does this sound like you? Is it written for the person who will actually read it?

A practical way to reduce the slop over time is a voice loop. After you edit an AI draft into something you would actually send, feed the final version back into the model and ask it to study the changes. Over time, memory, skills, or a voice profile can cut down the repeated edits and make future drafts sound more like you.

If your tools have memory or reusable instructions, this gets more valuable. The principle works even if they don’t.

Want a concrete version of this? Take 5-10 polished emails or posts you’ve already written, the ones you’d actually send, and ask the model to study how you write. Save that as your voice.

Better words, better output

The more domain vocabulary you have, the better your AI results get.

I noticed this with design. Learning even a small amount of design terminology helped me describe what I wanted more precisely. Screenshots, references, and better names give the model more context, especially in fields where you know the outcome you want but not how to describe it.

Brooks ran into a version of this using AI for motion graphics. He needed the model to help clarify the visual direction. You can’t prompt well for something you don’t have words for.

So pick one adjacent domain. Design, data, strategy, writing. Learn enough of the vocabulary to become a better collaborator.

Flip the Script

When you’re not sure how to prompt well, stop trying to provide the perfect instruction. Ask the AI to interview you first.

“Ask me questions before we proceed.”

“What do you need to know from me to succeed?”

I call this Flip the Script. Better answers come from better context. You already have the context. The model doesn’t. So let it pull the brief out of you.

This is the one tweak I want you to make this week.

Brooks has been saying the same thing about AI implementation in general. Don’t try to build a giant AI workflow all at once. One tweak a week.

He improved an existing podcast-prep workflow by adding one missing piece: AI-created task templates for OmniFocus. He noticed the friction first. Then he added Claude, screenshots, shortcuts, and an MCP-style integration around that one gap. He didn’t sit down and design a massive system.

Identify one repetitive friction point in your current workflow. Improve only that this week.

And for seven days, on every meaningful AI prompt, Flip the Script. Don’t try to write the perfect instruction. Ask the model what it needs from you first.

That’s it. One tweak. Seven days. Then see if the answers actually get better.