Deep work or parallel AI agents: what the research supports

AI agents now run tasks on their own while you do something else. Is focused, one-task-at-a-time deep work still the right default, or should you run several agents and juggle? We graded 19 research findings that people cite on both sides against the strongest independent tests published since. 6 held up: 3 of the 11 findings on focus and multitasking, and 3 of the 8 on working with AI.

Only one failed: that one person's performance drops once they supervise more than a few agents, and that grade rests on a single small test. The claim many people now work by, that running several AI agents in parallel makes a person more productive, has never been tested in a controlled study.

How we graded

We wrote down the 19 findings, our predicted grades and the grading rules before any search. We ran 205 queries in three literature databases and followed 23,054 citations of the original studies. Every test was extracted with the location of each number, and a separate review re-opened every source, checked each value and dropped the ones it could not verify.

Only strong tests by authors outside the original team count: meta-analyses of at least 10 effects, replications with at least 100 people (or 20 measured under both conditions), and field experiments with at least 200 randomized workers or tasks. 88 tests entered a grade. Replicated means independent teams found the effect again and Thin that one did. Not tested again is not a verdict that a finding is false.

Switching costs are real

Switching between tasks costs time: in a meta-analysis of 73 effects, people answered reliably slower right after a switch, and every qualifying test agreed. Doing two demanding things at once degrades at least one of them, in all 6 meta-analyses we could grade. And interruptions raise errors and the time it takes to resume: 10 of 11 independent tests found it. In one lab study, people made sequence errors on 7.3% of steps right after an interruption, against 1.5% without one.

The softer claims are mixed. Heavy media multitaskers are not reliably worse at filtering distractions once publication bias is corrected for. Batching email and notifications helped in 2 tests and not in 3.

The famous numbers mostly do not hold

We traced 12 popular statistics about multitasking and AI productivity to their sources. Only 1 holds: knowledge workers switch tasks about every 3 minutes, and an independent measurement confirmed it within a factor of two.

• "Switching costs up to 40% of productive time" is a remark on a web page, not a measurement.
• "23 minutes to get back on task" comes from a 2006 interview. The closest peer-reviewed figure, 25 minutes 26 seconds, is how long people took to return to an interrupted project while doing about two other things in between.
• "Only 2.5% of people can multitask" was measured in 200 students on one driving task and has not been tested again in a large study.

AI raises output, with caveats

The clearest AI result: generative AI raises output on standard job tasks. A meta-analysis of 1,161 estimates from 41 studies puts the gain at 16.7%, and 19 of 22 graded tests found a gain. A strict publication-bias correction shrinks it to 7.4%, still above zero.

Two other findings held up, one on a single test so far. Using AI on tasks outside what it does well lowers correctness. And reliable automation makes people check less: across 37 studies of partly automated driving, drivers paid less visual attention to the road.

Two popular claims are mixed. The gains are not reliably largest for less experienced workers: 5 tests found it and 2, both run by the tool's maker, did not. And the finding that experienced developers were 19% slower with AI tools, while believing they were 20% faster, came from 16 developers. Of three larger field tests since, one found experienced developers' own commits fell as they took on more review, and two found no slowdown. Across the AI findings, 14 of the 33 graded tests were run by the company that made the tool.

Parallel agents have not been tested

The closest evidence comes from operators of several robots or drones. In 4 small studies that located a limit, performance peaked or fell at 3 to 12 robots per operator, a median of 6.5. The one study large enough to grade compared 3 with 6 robots and found no drop. That is too thin to set a number.

So we used a simple model, which is not evidence. Each agent works alone for a while, then needs your attention, and every move to its thread costs you a switch. The number of agents you can keep busy is (run time + attention time + switch cost) divided by (attention time + switch cost).

Two vendor reports put the median turn of today's coding agents at 45 seconds and 63.4 seconds. At that length, one person can keep only 1 to 2 agents busy. If agents run 30 minutes and each check-in takes 2 minutes of attention plus a 2-minute switch, the limit is 8.5 agents, and half your supervision time goes to switching.

What this review cannot show

• Transfer. The evidence comes from lab tasks, students, coding, support and consulting work, and robot operators. None of it measures people supervising several AI agents.
• Scale. The replicated switch costs are milliseconds to seconds in lab tasks, while an agent check-in takes minutes, so they may be small next to review and waiting time.
• Fragile grades. Several grades sit one test away from another, and 10 of the 88 graded tests were checked against an abstract only.
• Job postings. We also planned to measure how employers' language about multitasking, deep work and managing AI agents changed in job postings from Ken's web archive. That panel was not completed, so it is not reported here.

What it means for how you work

• Treat each agent check-in as an interruption. Switching and interruption costs held up. In the model, checking 5 agents one at a time pulls you out of your own work 8.824 times an hour; checking all 5 in one block, 1.765 times.
• Lengthen agent tasks before you add agents. With turns of about a minute, the model says you can keep only 1 to 2 busy.
• Do not watch agents stream while you do your own deep work. Doing two demanding things at once degrades one of them.
• Review AI output as if someone else wrote it, especially when it has been right before. Complacency with reliable automation held up.

Deep work is still the right default for your own task. In the model, parallel agents pay off when each can work a long stretch without you and you check on them in batches.