Staying Sane When AI Agents Keep Changing Your Workflow

Every morning I sit down and try to work with agents, and every morning the same thing happens: while one agent is busy, my brain starts wandering toward the next thing it could be doing. Another area of the same repo. Another repo entirely. Another idea that looks harmless until I’m already paying the cost of switching.

That’s the part people underestimate. The agent itself is not the only task. The real tax is the constant opening and closing of context, plus the need to come back later and verify what happened. By the time you’ve done that a few times, your attention is already thinner than you thought.

The cost isn’t just time, it’s mental switching

Before agents, we already knew this feeling from chat, async coordination, alerts, meetings, and the usual modern noise. But agents make it more aggressive. They respond quickly, they stay verbose, and they invite you to keep going.

It feels like a stream that only ends when you stop feeding it. That changes the shape of work. You are no longer just doing one task; you are also managing a machine that will happily create more tasks if you let it.

That’s why I think of agents as a new organization full of very smart, very dumb, very high-maintenance coworkers. If you don’t put limits around them, they will destroy your sanity.

Why I had to stop letting the tool run my head

At some point, the problem stopped being theoretical. The interactivity started to feel addictive. The dopamine hit from AI tooling was strong enough that I kept thinking about using it for almost anything, and the FOMO around “not creating value” got real.

Social media made it worse, because everywhere you look there is another person showing off a new fleet of agents they’ve set loose on their work. Maybe that future exists for some people. For me, the pace felt unhealthy.

I had a moment where I was prompting until 2:00 AM, and the next day I still felt like I was just getting started. After a day of prompting for more than 14 hours, I felt like a bot. I was so locked into that pace that I was almost expecting family and friends to reply at machine speed too.

That was the point where I stopped romanticizing it. The rule became simple: I should control the tool, not let the tool control me.

The boundaries that keep me usable

The fixes are not glamorous. They’re the old techniques, which is probably why they work.

• I use timeboxes.
• I keep limited context windows.
• I schedule specific hours for agent work.
• I leave room for undivided attention.
• I sometimes avoid agents entirely and think on paper, handwriting, or walking without my phone.

I also pay attention to how I feel after a day of interacting with agents every hour. If I’m getting too scattered, that matters. If I’m feeling pulled into the machine, that matters too.

I now stop coding-related prompting at 21:00 so I have a few hours of rest before bed. That one boundary matters more than it sounds like it should.

Four windows, not infinite ones

One thing I found useful is having a maximum number of concurrent agent windows. Right now it’s four. Each one is in planning or executing mode, and I can move between them without losing too much of the thread.

Before AI, I had two work-in-progress windows. So yes, my focus capacity has roughly doubled in one narrow sense. But that does not mean my attention is infinite. It just means I can manage a little more without immediately collapsing.

That still comes with a warning: there is a difference between handling four active threads and letting four agents multiply into twenty unplanned thoughts. The second version is where the damage starts.

The notebook comes before the prompt

One of the best guardrails I added was going back to basics every morning.

Before I write a single prompt, I start with a notebook and a pen. I dump what I want to achieve. Then I timebox the day around that. The structure matters because otherwise the agents will happily eat the day whole.

I keep a small concrete todo list and I actually track it. I can tell when it’s working because I’m marking off more items. I also make sure some of those items are non-digital: watering plants, cleaning the 3D printer, reading, anything that reminds me I have a life outside the prompt window.

That mix is important. If everything becomes agent-shaped, everything starts to feel urgent and abstract.

When a small task turns into a monster

The clearest example is the blog post that was supposed to be simple.

The todo item was basically: write one small blog post about AI. That was too vague, and vague tasks are where agents become dangerous. Instead of writing for an hour, I spawned an agent to refactor the blog from Gatsby to Astro. Then I started choosing templates, making design decisions, fixing the build, and chasing trial-and-error changes.

I spent six hours on the site and wrote no post.

The next day I changed the rule. The task was no longer “work on the blog.” It was “write the blog post.” That day I spent about two hours writing, which was the actual thing I had said I wanted to do. Later, I allowed some side projects, but only because they were planned and timeboxed.

That’s the pattern: support work is fine when it supports the main thing. It’s not fine when it becomes the main thing by accident.

The 80/20 rule I keep coming back to

My rough rule is simple: 80% on the writing, 20% on whatever makes future writing easier, faster, or better.

That 20% is where the useful experiments live. It gives you room to improve the system without letting the system become the project. But the 80% has to stay dominant, or you end up polishing the machine while the work you came to do never happens.

Even after years of trying to be disciplined, this is still a constant struggle.

The question I ask before I detour

I’ve also learned to ask a more annoying question: not just whether a side task could help, but how it helps the main task.

Sometimes the answer is yes. Sometimes the answer is that the thing is interesting but not necessary. And sometimes the answer is that I need to force friction on myself so the decision is harder to make on impulse.

That friction can be as simple as saying, “This gets 30 minutes.” If I know I only have that much time, I’m more likely to reconsider before I start. It’s a small barrier, but it creates just enough pause for me to notice that I’m about to leave the main road.

Timeboxing is not a cure, it’s a seatbelt

Timeboxing does not magically solve the attention problem. It just makes the damage less likely.

If I let myself, I can still fall into a rabbit hole. I once asked GPT for deep research on companies that allow vibecoding in interviews while I was writing about AI and companies not adapting. It came back with a 10-page essay. I read it for about 20 minutes, learned something useful, and then lost the thread for the day.

That’s the tradeoff. The tool can absolutely teach you things. It can also derail you so completely that the main work disappears.

Keeping the tool useful without letting it become the weather

I don’t think the goal is to reject agents. I default to them now. I just have to make room for the parts of life that keep me sane: reading, handwriting, walking, family, friends, and enough unoptimized time to think clearly.

The healthy version of this work is not infinite productivity. It’s being able to use the tool without becoming dependent on its pace.

That means protecting attention on purpose. It means putting guardrails on the day before the day gets away from you. And it means remembering, over and over, that the point is not to keep the agent busy.

The point is to keep yourself working on what matters.