AI fatigue: what happens when product teams can't keep up with their own agents

For the past two years, the conversation around AI in software engineering has focused on one thing: productivity. Engineers are shipping faster, writing more code, and completing work in hours that once took days. Every new model promises another leap forward.

What gets far less attention is what all that speed demands from the people using it. Guillaume Moigneu, Field CTO at Upsun, has spent the past year watching engineering teams adapt to AI-assisted development. For experienced architects, the change has been transformative. "I can still practice architecture, driving the agents with specifications, and then the AI actually gets into the nitty-gritty parts, writing the code," he says.

But those gains look very different once AI moves beyond individual developers and into team-wide workflows.

The fatigue nobody is naming

AI fatigue has become a common complaint, but it's often misdiagnosed. Most people assume it comes from tool overload or relentless hype. Guillaume sees something different: the cognitive burden of supervising multiple AI agents at once.

"I started to actually have multiple agents at the same time on my laptop," he says. "I was doing four, five different things, maybe on the same project, maybe on different projects, and I was getting a bit overwhelmed by all this." The problem wasn't the amount of work; it was the constant supervision and context-switching. Always one prompt away from wrapping this up. 

Most developers rely on long stretches of uninterrupted focus to solve complex problems. AI agents don't operate on that schedule. They finish work whenever they're ready, send back results at unpredictable moments, and expect immediate review. Running four or five agents at once means being interrupted four or five times by code you didn't write but are still responsible for approving.

"Every time you get into something, then you switch, you reswitch, you get back, and you get lost," Guillaume says. "It felt really overwhelming." Our brains also need rest. Most studies show that our brains follow an Ultradian cycle where they can perform for 90 minutes and then need some rest. Coding agents were creating an infinite strain without breaks.
 

The trust problem underneath

The fatigue is not just about attention; the bigger issue is responsibility. When an agent writes code, the engineer is still accountable for what it produces. That accountability does not scale the way the output does.

"I have some trust issues with it," Guillaume says, "because there is so much code being produced that I'm not even sure that all that code is correct or even really does what I want it to do." 

The review burden grows with every agent added to the workflow. Reviewing AI-generated code isn't the same as reviewing a colleague's work. Human reviewers infer intent from context and conversation. AI has no intent to clarify. Its output can be syntactically correct, pass every test, and still solve the wrong problem because it faithfully executed the prompt it received, rather than the one the engineer meant to give.

Some teams have responded by adding more process on top: more checkpoints, more reviewers, more gates. Others have tried to automate the review layer itself. Neither fully solves the problem because the issue is structural. The tools were built for individual productivity. The workflows surrounding them were not redesigned to match.
 

When the team feels it

What starts as an individual productivity gain can quickly become a team bottleneck.

Guillaume describes a dynamic he has seen play out repeatedly: one engineer, highly proficient with agents, produces code at a rate the rest of the team cannot absorb. "If you had one engineer like this, that would create a lot of issues for the others," he says. The fast engineer isn't doing anything wrong. But their output quickly exceeds the team's capacity to review, integrate, and build on it. Bottlenecks shift from writing code to reviewing and validating it,

"What started as a tooling gap hardens into resentment between teams, exacerbating already existing silos," Guillaume says, "and that's much harder to repair."
 

What this is actually telling us

AI fatigue, properly understood, is not a signal that the tools are too powerful or that adoption has gone too far. It is a signal that the infrastructure around the tools has not kept pace. The models have improved faster than the workflows. Individual capability has outrun team coordination.

Guillaume offers a simple analogy: "The reason you want to buy a new car is that you have an issue with your existing one or because you don't have one yet." In other words, teams rarely change processes unless the current ones stop working. AI fatigue is that signal. It tells you that existing workflows no longer fit the way modern teams build software.

The question for engineering leaders isn't whether to slow AI adoption. It is about redesigning how teams work so humans spend less time supervising agents and more time solving problems. That's the work most teams haven't started yet - and it is exactly what Upsun Dispatch™  is built to support.