How to Run Multiple Projects With AI Agents Without Burnout

Most advice about AI agents is about getting more done. This article is about a different problem: what happens to the person in the middle when more gets done.

If you run several things at once (a client, two side projects, an app that half works, a job hunt), you already know the feeling. The tools get faster and you get more tired. That is not a coincidence, and it is not a character flaw. It is a design problem, and design problems can be fixed with some plain rules.

Here are the rules I use. None of them needs special software.

The limit is your attention, not your tools

Ideas are not scarce. Most builders have a notes app full of them and a few domain names bought at midnight. What is scarce is the single narrow pipe that every idea, client and invoice has to pass through: your attention.

The bottleneck is not a flaw in the system. It is the system's most expensive component, and you are it.

Adding AI tools does not widen that pipe. Each new tool is one more thing asking to be looked at. So the first job is to decide what gets through at all.

There are four things you can do with a task, and only one of them involves you doing it:

• Delete it. Most tasks survive because nobody asked whether they should.
• Automate it. If it happens the same way every time, a script or a scheduled job can own it.
• Delegate it. If it needs judgement but not yours, hand it to an agent with a clear brief.
• Do it. Whatever is left, and it should be the smallest pile.

Ask in that order. People usually ask in the reverse order, which is how they end up doing everything.

It also helps to be plain about what an agent is. Strip away the marketing and an AI agent in 2026 is a language model given a goal, some tools and permission to loop: try something, look at the result, try again. It has no ambition and no secret plan. It is closer to an intern who never sleeps. Useful, quick, and in need of structure.

Give every project one owner

Put two people in charge of something and you have built a polite machine for nobody doing it. Two agents in charge of something is the same machine, only faster and with logs.

A thing that belongs to everyone is maintained by weather.

The rule in my own system is boring and absolute: every project that matters has exactly one accountable agent. One owns the fleet of small apps. One owns media. One owns spend. One owns the job hunt. Each can borrow help, but none of them can say "I thought someone else had it." There is no someone else.

An agent, in this sense, is less a personality than a contract. Mine is a small file that says what the agent is for, what it takes in, what it hands back, and what proof it leaves behind when it claims to have finished. The last field is the one most often left empty, and it is the one that matters most.

Two more fences are worth building early:

1. A memory the agents can read but not change. I keep my goals, decisions and notes in a private repository. Agents may read it whenever they need to. They may not create, move or delete a single file in it, however much they would like to tidy it.
2. Permissions written into each project. What an agent may do should be on the page, in plain words, where every agent loads it.

Brief the outcome, then ask for a receipt

The worst brief I ever gave an agent was a list of steps. Open this file, change that line, run this command, commit. It did exactly as asked. The steps were wrong in a way I only noticed afterwards, and the agent, having been told how and not why, had no reason to notice either. It reported success. Nothing I wanted had happened.

A better brief is three sentences:

• Done means: the result you want, stated so it can be checked.
• Proof is: what the agent must show you. A live URL, a log, a commit hash.
• Do not: the things it must leave alone.

Then hold the line on the second sentence. Somewhere in your backlog is an item marked done that is not done. You do not know which one. The fix is to stop taking anyone's word for it, your own included.

Done is a claim. Evidence is the receipt.

In practice, every finished item carries a link or a path that a person or a machine can check. An item without one is not done. It is reported done, which is a different status and should look different on your board.

The same idea applies to trust. "I trust it" is a feeling, and feelings are poor infrastructure. If an agent has been promoted to working alone because it has been fine lately, check what "lately" means. Often it means you have not looked. Count the runs, count the ones that came with evidence, count the ones you had to fix, and let those numbers decide how much freedom it gets.

Read one report and stop checking

Burnout in this kind of work rarely comes from the work. It comes from checking. Ten things are running, any of them might be broken, and so you look. Then you look again.

The cure is one report at a fixed time. Mine arrives at half past seven every morning. I do not write it. A machine that stays switched on gathers the state of everything overnight, an orchestrating agent summarises it, and it lands where I will see it with the first coffee. It has the same fixed sections every day, so I can read it half awake.

Two details make this work:

• Unknown is allowed. Where a number is not known, the report says Unknown. Not zero, not a hopeful estimate. It is the most honest word on the dashboard.
• The report is the checking. Having read it, you are entitled to not look again until tomorrow. That is the whole point.

Things will still break. A dependency updates overnight. A scheduled run is cancelled because a machine went to sleep. The first rule of incidents is to slow down, which sounds backwards. The instinct is to try six fixes at once from a phone at the bus stop, and that produces a second incident on top of the first. Notice, pause, fix one thing, write down what happened so the system is a little harder to break next time.

Build a week that does not ask how you feel

I used to plan weeks the way some people pack for holidays: everything, just in case. Twenty items on Monday, four done by Friday, sixteen carried forward with a faint smell of failure. The problem was not the doing. It was the packing.

The board I run now has five slots. Five outcomes for the week, each sized so it can be finished and evidenced by Friday. "Work on the affiliate feature" is not an outcome. "Affiliate links live on the travel app, with one click recorded" is. One of those can be completed. The other can only be continued.

Friday has a second job: kill one thing. A side project you enjoy explaining at parties, an agent with a lovely name that has done nothing since June. A system that only grows will, in time, only maintain.

Rest belongs on the same board. I do not resent the small computer on my desk for needing power, cooling and the odd restart, and I would not be proud of running it hot. If its fans screamed every day I would call that a fault. The operator deserves the same maintenance as the hardware.

None of this depends on motivation, which is weather: bright some mornings, drizzling on others. Rhythm is climate. The report arrives, Monday gets five outcomes, Friday loses one darling, and none of it requires you to feel anything in particular.

Where to start

You do not need all of this at once. A sensible order:

1. Sort this week's tasks into delete, automate, delegate and do, in that order.
2. Give each project that matters one named owner.
3. Rewrite your next brief as done means, proof is, do not.
4. Refuse to mark anything done without a link.
5. Set up one daily report and stop checking between reports.
6. Cap the week at five outcomes and kill one thing on Friday.

Install one habit, let it settle, then add the next.

These ideas come from The Unbothered Machine: How one person runs a hundred apps with AI agents (and stays calm about it), the operating manual I wrote from running this system at one desk. It has 100 short chapters, each with its own diagram, and comes with the Machine Kit of 10 templates, including the agent contract and the daily status. It is £9 on Gumroad with a 30-day refund: matsiems.gumroad.com/l/unbothered-machine. You can read three chapters free first at matsiems.com/books/the-unbothered-machine.