Two ideas have been treated as settled truth in knowledge work for the better part of two decades. First, that multitasking is a productivity killer — that splitting attention across parallel tasks costs more than it gains. Second, that Agile is built on single-piece flow — limiting work in progress, finishing one thing before starting the next, and treating context-switching as waste to be eliminated. These two ideas reinforce each other, and they’ve shaped how nearly every modern software and product team organises itself.
Here’s the hypothesis I want to put forward: the arrival of capable AI agents partially breaks both assumptions. Not because the underlying research was wrong — it wasn’t — but because that research measured a specific actor with specific limits: the human brain. When the thing doing the parallel work is no longer a human, the cost structure that justified single-piece flow changes fundamentally. This article lays out the argument, and — because a good hypothesis has to survive its own counter-evidence — the significant limits on it.
Why Agile Rejected Multitasking in the First Place
The case against multitasking in Agile isn’t dogma — it rests on genuinely robust evidence about how human cognition works. The core mechanism is context-switching cost. Every time a person stops one task, unloads its mental context, loads another, and tries to resume, they pay a tax in time and accuracy.
The numbers are stark. Research reported by Harvard Business Review found the average digital worker toggles between applications and websites roughly 1,200 times per day — about one switch every 24 seconds across an eight-hour day. The American Psychological Association’s research indicates chronic multitasking and frequent context-switching can consume up to 40% of a person’s productive time. Microsoft researchers found the typical knowledge worker spends less than three minutes on a screen before switching. The result isn’t true parallelism — it’s what one analysis aptly called “serial theft of attention.”
This is precisely why Agile — and Kanban in particular — enshrined Work In Progress (WIP) limits. As the Kanban literature puts it plainly, WIP limits encourage focus and reduce multitasking; context-switching between work items carries cognitive costs that compound as the number of active items increases. A classic worked example from the WIP-limit literature: three sequential projects take an average of 60 days each to complete, versus 89 days each if multitasked — a 30%+ improvement in time-to-market from not multitasking, before you even count the real-world context-switching penalty. The message was unambiguous: stop starting, start finishing.
These principles are baked into how I’ve always described good governance in my own work, including Part 4 of my UI/UX Playbook. The evidence was, and remains, sound — for humans.
The Hidden Assumption: The Worker Is Human
Here’s the crucial observation. Every piece of evidence above measures the same actor: a human being with a single, serial attention system. The cost of multitasking is not a law of physics. It’s a property of human cognition — specifically, our inability to hold multiple rich task-contexts active at once and switch between them without penalty.
Agile’s single-piece flow, then, was never really a statement about the nature of work. It was a statement about the nature of the worker. WIP limits are, at root, a workaround for a hardware limitation in the human brain. And that reframing is what opens the door to the hypothesis: if you change the worker, you change the constraint.
What AI Agents Change
AI agents — autonomous systems that pursue a goal across multiple steps rather than answering a single prompt — don’t share the human attention architecture. Several agents can run genuinely in parallel, each holding its own full task-context, none paying a switching penalty to exist alongside the others. Where a human “doing three things at once” is really one attention system thrashing between three contexts, three agents doing three things is genuine simultaneity.
This is already visible in how modern agentic work is being structured. The professional’s role increasingly shifts toward direction, quality review, and refinement rather than production from scratch — a human orchestrating multiple parallel workstreams that agents actually execute. The cognitive overhead of maintaining situational awareness across parallel workstreams, historically a persistent drain on human focus, is substantially reduced when an agentic layer tracks progress, surfaces blockers, and reports back.
If that holds, it genuinely does overturn the two starting assumptions — with an important restatement. Multitasking isn’t suddenly good for humans; rather, a human can now orchestrate parallel work without personally performing it in parallel. And Agile’s single-piece flow stops being the universally optimal pattern, because the actor it was designed around is no longer the only actor on the team. I’ve explored the mechanics of how this orchestration works in practice in the article on agentic AI workflows.
The Counter-Evidence: Where the Hypothesis Breaks Down
A hypothesis worth publishing has to state the conditions under which it’s wrong. And there’s a genuinely important finding that constrains this one sharply.
Research from Google in 2026 found that multi-agent coordination improves performance by 81% on parallelizable tasks — and degrades it by up to 70% on sequential ones. This is the single most important number in this whole discussion. Parallel execution is not universally superior; it is superior only when the work is genuinely parallelizable. When tasks have dependencies — when task B needs the output of task A — running them in parallel doesn’t just fail to help, it actively harms the outcome.
Notice what this means: the old Agile wisdom doesn’t vanish — it relocates. The discipline of sequencing dependent work, of not starting what can’t yet be finished, remains exactly as valid for agents as it was for humans on truly sequential tasks. What’s changed is that a whole category of work previously forced into sequence by human limitation can now be genuinely parallelised. The skill shifts from “limit WIP” to “correctly classify which work is parallelizable and which is sequential” — and the penalty for misclassifying is severe in both directions.
There are further constraints. Parallel agents introduce coordination overhead: a multi-agent pipeline can consume roughly three times the tokens of a single-agent approach, and each added agent introduces measurable coordination latency. And they introduce a subtler failure mode — context fragmentation, where parallel agents each operate on a slightly different slice of organisational knowledge and quietly diverge. Quality collapses not from the parallelism itself but from agents no longer sharing the same understanding.
A Revised Synthesis
So where does this leave the two original assumptions? Neither is simply “true” or “false” any more. A more precise formulation:
- “Multitasking is always bad” is now false as stated — but only because the multitasking has been offloaded to actors that don’t pay the human context-switching tax. The human still shouldn’t multitask; they should orchestrate.
- “Agile single-piece flow is universally optimal” is now false as stated — but its underlying logic survives intact for dependent, sequential work, which the Google finding shows is actively harmed by forced parallelism.
- The new core skill is dependency classification. The value has moved from limiting work-in-progress to correctly separating parallelizable work (hand to parallel agents) from sequential work (keep in disciplined flow).
- A shared context layer becomes non-negotiable. Parallel agents only stay coherent if they draw on the same institutional knowledge — the modern equivalent of a co-located Agile team’s shared understanding.
What This Means for How Teams Should Work
If the hypothesis holds even partially, the practical implications for product and engineering teams are significant.
- Rethink WIP limits for the human-plus-agent team. A WIP limit on a human’s personal focus still makes sense. A WIP limit on the parallel agent workstreams that human is directing may be leaving enormous throughput on the table.
- Invest in dependency mapping. Since misclassifying sequential work as parallel carries a heavy penalty, the ability to see and map task dependencies clearly becomes a first-order team capability.
- Design the orchestration interface deliberately. A human directing several parallel agents needs an interface built for exactly that — a challenge I’ve written about directly in designing for AI agents.
- Keep Agile’s values, update its mechanics. The Agile emphasis on fast feedback, iteration, and finishing work still holds. What’s up for revision is the specific practice of serialising everything — a practice built for a constraint that no longer fully applies.
Closing: A Hypothesis, Not a Verdict
I’ve framed this deliberately as a hypothesis, because it deserves that humility. We are early, the tooling is immature, and the Google finding is a strong reminder that parallelism has hard limits rooted in the structure of the work itself, not just the worker. But the core claim seems robust: Agile’s rejection of multitasking was a rational response to a human cognitive constraint, and AI agents relax that specific constraint for the subset of work that is genuinely parallelizable.
The teams that win in this transition won’t be the ones that abandon Agile, nor the ones that cling to single-piece flow as an article of faith. They’ll be the ones that understand exactly why the old rules existed — and are therefore able to tell precisely when a new class of worker makes those rules safe to break.
This is an opinion piece presenting a hypothesis for discussion, not an empirical claim. I’d genuinely welcome disagreement and counter-argument.
Related reading: Agentic AI Workflows: The Invisible Glue of Modern Business Operations · Designing for AI Agents · UI/UX Playbook Part 4: UX Governance
Rethinking how your team works in an agentic world? Get in touch.