"

AI: The illusion of productivity - The coordination overhead problem

In the previous article of this series, AI: The Illusion of Productivity, we explored the flow-state problem: reviewing AI-generated code demands more focused attention than writing it, and the rhythm of a chat-based interaction makes that attention hard to sustain. This time, we turn to what happens when that friction gets multiplied. Running several agents in parallel doesn't just add more code to review, it changes the job itself, from writing software to coordinating it.

***

If a single agent already disrupts the conditions good work needs, running several in parallel compounds every failure described so far and adds a few of its own. A multi-agent setup multiplies cognitive demand before it multiplies output. Whether output follows depends on whether you have any cognitive capacity left over once the coordination bill comes due.

Working memory has a hard ceiling: this is the foundational finding of John Sweller's cognitive load theory. Push past it, and performance doesn't gently degrade. It collapses: accuracy drops, creative reasoning stalls, errors compound. Managing three agents, each producing output that needs contextual evaluation, exceeds that ceiling faster than most of us expect, because we're used to thinking of "capacity" in terms of time available, not attention available.

It gets worse with each switch, not just with each additional agent. Sophie Leroy's 2009 research on attention residue found that when people move from one task to another, part of their attention doesn't come with them: it stays behind, clinging to the unfinished task, quietly consuming processing capacity in the background. Three agents isn't three times the cognitive load of one. It's more, because every switch between them generates residue that compounds against the next.

There's now a reasonably large empirical signal pointing the same direction. ActivTrak's 2026 State of the Workplace report, based on 443 million hours of behavioural data across more than a thousand organisations, found the average focused, uninterrupted work session has fallen 9%, a continuation of a three-year downward trend, at the same time AI adoption surged to 80% of the workforce measured. More AI tools in daily use, less deep focus available. And a BCG survey of nearly 1,500 full-time employees found a meaningful share reporting genuine acute cognitive fatigue tied specifically to managing multiple AI systems at once: mental fog, slower decision-making, the felt sense that thinking itself had become crowded. People have started calling it "AI brain fry" and the name is doing real descriptive work.

Underneath all of this sits a role change that rarely gets named honestly. When you're coordinating several agents, you're no longer an engineer. You're a manager, at best a reviewer, at scale, and neither role comes with the cognitive conditions that make it possible to do well. That's not a promotion. Management, done properly, is its own discipline: deciding which of several competing streams of work deserves your attention right now, holding enough context on each to make that call correctly, and constantly re-prioritising as new information arrives from all of them at once. None of that is free, and none of it is the same skill as writing code. Doing it across three simultaneous, independent problem contexts isn't delegation. It's supervision with extra steps.

It also assumes the tasks being spread across those agents are the easy kind:  the kind AI handles well. They usually aren't, for exactly the people most likely to be running several agents at once. METR's own study, and related research, confirms that AI coding tools show dramatically different results depending on the terrain: strong gains on greenfield, well-scoped, isolated tasks; weak or negative results on complex, mature codebases where deep contextual knowledge is the actual bottleneck. The senior developers and tech leads reaching for the multi-agent setup are, more often than not, the very people working in the second kind of codebase.

This is where the second friction from my own experiment belongs: the reviewer-and-implementer loop that never converged. Two roles, both nominally "agents," passing work back and forth without reaching resolution. It was a small, self-contained demonstration of exactly this dynamic: a second layer introduced to improve quality instead produced a tie that only I could break, and I could only break it by staying close enough to the actual details to notice the loop wasn't closing. The supervisory burden that created was the precise opposite of delegation. Of everything in this piece, it's the clearest personal evidence I have that multi-agent setups redistribute cognitive load rather than removing it.

There's a particular seduction in this for senior developers and tech leads specifically: the fantasy of operating above the code, conducting rather than typing. I understand the appeal; I felt it myself, briefly, before the coordination costs caught up with me. But "above the code," on problems that genuinely require code-level understanding, isn't leverage. It's distance. And distance from a problem you're accountable for is not a good place to be standing.

All three of the problems above share a hidden assumption: that building faster is the right thing to be optimising in the first place. I don't think it usually is.

***

This is the third article in the AI: The Illusion of Productivity series. Part four comes out on September 22: we'll explore what actually slows software down. In most organisations, the constraint on shipping good software isn't how fast code gets written. It's upstream: knowing what to build, why, for whom, and whether the architecture can support the answer. AI that speeds up typing doesn't fix that, it produces more output against unclear specs and under-designed architectures, and makes the real bottleneck harder to see.

***

We wrote this article with the support of AI. If you’re interested in how we use these tools as part of our writing process, you can read more here.

***

Linkography

Cover image:
Photo by Numan Ali on Unsplash

Related posts