Recent months have been crazy. AI changed software engineering and there is no way to escape it. Productivity growing exponentially, small teams beating huge companies, new models emerging every day. It’s a major shift in the way we work.

Unfortunately, it brings a lot of challenges. Experts on X telling you you’re being left behind if you’re not running 10 sessions at once. Code quality dropping to the ground. Constant exhaustion from context switching. It’s a clear path to burnout. I’ve seen it myself - on me and on my team.

But it doesn’t have to be this way. Over the last few months I’ve adjusted how I work with AI - how many agents I run, when I look at them, what I let interrupt me - and the chaos got manageable.

Don’t go in at the deep end

Let’s say it out loud: it’s ok to work on one thing at a time. AI makes you way more productive than ever even without running multiple agents at once.

The fastest way to burn out is to jump straight to ten sessions before you can run one well. You end up babysitting all of them, reviewing code you don’t fully understand, and feeling busier than ever while quality slowly drops.

Before scaling up, learn how to work with a single agent efficiently - without constantly interfering or guiding it. Your agent should be able to take a task from start to finish on its own. There are many ways to achieve this: proper CLAUDE.md/AGENTS.md files, clear prompts, related skills, linters, plan mode and so on. The single highest-leverage habit is the simplest one: every time your agent makes a mistake, add a rule that prevents it from happening again. That’s how the agent gets trustworthy over time.

If you feel like you’re wasting time waiting for the AI to finish, try fast mode (available in Codex for 2x tokens or in Claude as extra usage). But fix the autonomy problem first - going faster doesn’t help if you’re still in the loop on every decision.

Scaling up

Once your agent can run a task end-to-end without you, you’re ready to add more. Four things made the difference for me: scaling slowly, planning my day in advance, killing the noise, and using a real orchestrator.

Scale slowly

Don’t jump from one session to ten. Start with two. Then maybe three. If you can keep them running without losing track of what either is doing, add another. If you can’t, drop back for a few more days. The right number is personal - some people max out at three, others run eight. Find yours, don’t copy someone else’s.

Plan your switches

Constant context switching is a real problem - and it’s nothing new. Before AI, I hit it as I took on more responsibilities: reviewing CVs, conducting interviews, code reviews, finishing my own tasks. What pulled me out was time-blocking my day in advance, and the same principle works for AI sessions.

I start all my sessions at the same time, and I make sure each agent has everything it needs to run end-to-end. Then I don’t look at them - I move on to whatever’s next on my list: reviews, emails, my own coding. I don’t care if a session fails or finishes early. It can wait. It’s not going to starve. I check them when it’s their turn in my plan, not when they ping me.

Reduce the noise

AI didn’t invent notification overload, but it cranked it up by 100x. Every agent wants to ping you the moment it’s done. Every PR comment, every CI failure, every Slack message competes for the same attention. Don’t turn async into sync.

The fix is mechanical: turn the notifications off. Agent pings, PR comments, channel messages - anything that isn’t urgent. The only thing allowed to interrupt me is production on fire.

Here’s a dumb analogy: you’re chopping vegetables for soup. Someone asks you to also peel a potato. You don’t stop mid-cucumber - unless the kitchen is on fire, the potato can wait.

Use proper tools

Running multiple agents from separate terminal windows works for two or three sessions, but it falls apart fast. You need an orchestration layer - one place to start tasks, watch progress, and review diffs without juggling windows.

Codex and Claude both ship one. If you don’t want to be tied to a single provider, Conductor and Emdash are both good. Try a couple, pick the one that fits how you actually work.

It’s a skill, not a setting

Working with AI well isn’t just about picking the right model or writing the perfect prompt. It’s about managing yourself. The tools got faster, but your attention budget hasn’t changed.

If you don’t protect your focus, AI won’t just speed you up - it’ll speed up the chaos too.