For Deep Work, Don't Run AI Agents In Parallel
Parallel agents can clear the labour, but the judgment still queues with one person.
By Gabriel Tan | July 2026
Last month I tried to build three pitch decks at the same time.
Three prospects, three different stories, each with its own code name so I could keep them apart. I used Cowork to run the research on all three at once. It felt efficient. Three agents working in the background while I got on with other things.
To be honest, it turned into a mess.
The agents kept coming back to ask for my judgment. One needed a call on positioning. One wanted me to pick which angle to lead with. One asked which case study to feature. Three different pitches, three sets of questions, all landing at the same time. I started mixing them up. The code names that were meant to keep them separate just made me more confused about which was which.
When all three decks were finally done, none of them were any good. Not one of the three. I deleted all of them and started again, one deck at a time.
The tool removes the labour, not the judgment
Here is what went wrong, from first principles.
An AI agent is very good at taking labour off your plate. Searching, pulling data, drafting a first version, formatting the output. What it cannot take off your plate is the judgment. Every time a real decision comes up, the work stops and waits for you.
That is fine when there is one stream. When there are three, all three stop and wait for the same person, and that person is you.
This is the part the "run more agents" advice misses. Running agents in parallel does not multiply your attention. It strips the labour out of three jobs and then queues all three decisions at the one place that cannot be cloned, which is your head. The agents run in parallel. The judgment does not. The queue backs up at you, and the work moves only as fast as you can clear it.
So the real question is not how many agents you can run. It is how much of the work is labour and how much is judgment.
Split the labour from the judgment, not your tasks
Most work is a mix of both, so do not sort your tasks into two piles. Sort the parts inside each task.
Take the labour parts and run them in parallel. Research, transcription, first drafts, pulling the same numbers every morning, formatting a report to house style. You already know what good looks like before the agent starts, so your job is review, not decision. Run five of these side by side and check them in one pass.
Take the judgment parts and do them one at a time. The angle of a media pitch. The order of proof in a positioning deck. What to cut from a crisis statement so the point survives. Each of those is a decision only you can make, and decisions do not split three ways.
My pitch decks are the proof. The research ran fine in parallel. The deck did not, because a deck is a chain of decisions, and I had given each one a third of my attention and asked it to pass as my full thinking.
The sort test is simple. If you already know what good looks like before the agent starts, it is labour, so parallelise it. If the agent has to stop and ask you what good looks like, that is judgment, so protect it.
Protect the wait
There is one more habit worth breaking, and it is the one that cost me the most.
When a single agent is running on the judgment work, the wait feels like dead time. The urge is to fill it by starting a second deck. Do not.
Stay on the same problem. Read what came back. Work out the next decision before the agent has to ask for it. The wait is not lost time. It is the cleanest block of thinking time you will get on that exact problem, and starting a second job throws it away just to feel busy.
What to do this week
Pick your next real piece of judgment work. A pitch, a strategy, a difficult client note.
Spend fifteen minutes marking which parts are labour and which parts are judgment. Hand the labour to your agents and let them run together. Keep the judgment for yourself and work it one decision at a time, start to finish, before you pick up the next thing.
This is the same call you already make with your team. You hand off the labour and you keep the decisions that carry your name. Agents do not change that rule. They just make it easier to forget, because they make it so cheap to start three things at once.
The rule did not change, only the tools
We can run more work in parallel than ever, and that is worth using.
But you still cannot outsource the thinking, and you still cannot think about three hard things at once. That was true with a team of people and it is true with a team of agents. The tools got faster. Your one head did not.
The skill now is not running as many agents as you can. It is knowing which part of the work should never be run that way.
Which piece of judgment work are you running in parallel right now that deserves your full attention, one decision at a time?
Gabriel Tan is the founder of Mekong Bridge Advisory. He builds structured execution systems for PR and communications firms.
