I've always been good at multitasking. Context-switching, juggling projects, keeping several threads alive at once – it's just how I work. So when agentic AI tools started becoming genuinely useful, I figured it would be a natural fit. I was right, kind of. But what I wasn't prepared for was how tired I would be at the end of the day.
Over the past year, my ways of working have changed faster than ever in my career. While this shift has made some things easier, it hasn't been a cakewalk. Something about working alongside AI agents – particularly the agentic, multi-step kind – has proven to be cognitively heavier than I expected.
What could be the root cause? Current agents are still relatively slow. And that latency, paradoxically, invites multitasking. While the agent works, you start something else. Then it finishes and needs your input. You switch back to the original context. Then it needs something again; and now you've lost the thread of the other thing. Your attention fragments. In terms of raw output, you get more done – but leave the office feeling like you've been pinballed between myriad to-do lists.
Turns out I'm not alone. Recently, I came across an article aptly titled The Agent Paradox that helped me identify what I'd been feeling. The author, Mark Llewellyn Dyer cites a study that shows that modern knowledge workers switch between apps with increasing frequency. In 2004, this would be every 2.5 minutes – and by 2023 the number had dropped to mere 47 seconds. Layer agentic workflows on top of that, and you've got a recipe for burnout dressed up as productivity.
The study suggests that people are actually quite bad at monitoring automated systems that occasionally require fast, accurate responses. Which is precisely the workflow that dealing with AI agents requires. This is essentially a design problem. And somewhat embarrassingly – given that solving design problems is literally what we do at Taiste – it took me a while to see it that way.
Inspired by this reframing, I've started making some practical changes. Here are a few things that have actually helped:
Do some of it yourself. AI handles routine work brilliantly, but for novel problems or complex debugging, I've found it often goes off track. Using your own intelligence on hard problems is often the right tool for the job.
Group similar tasks. I work across architecture, full-stack development, database optimisation, production incidents – across multiple clients and stacks. Jumping between all of these is cognitively expensive. Batching similar work, even loosely, makes transitions feel less demanding.
Be patient with yourself. Technology is moving faster than humans adapt. If things feel overwhelming or even a bit hopeless sometimes, that's a reasonable response. Change takes time.
The deeper question I keep coming back to is this: if an agentic system is technically efficient by every conventional metric, but the person on the other side of the screen is burning out using it – where exactly does the fault lie? Designing these systems with humans in mind should be a high priority, and we're only at the beginning of figuring this out.
