Burnout by the Kindness of Agents
AI might gain total control and destroy humanity one day, but I worry it’ll burn me out far earlier than that.
When I read the online discourse about AI-native work, it sounds like commentary on extreme sports performed by superhumans. I get the impression everyone is running swarms of agents overnight burning billions of tokens per month, building who knows what, and I often come across wild statements like someone describing a $1,000 per day of token spend as reasonable. Meanwhile, I get brain-fried when I run three agents in parallel.
In my world, when an agent identifies a bug and "solves it," that's only the beginning. I then have to load the necessary technical and business context for that part of the product into my head, review the code to ensure the agent didn't create needless or breaking changes, confirm whether the tests it built are testing the right things, and ensure its overconfidence didn't guide it to refactor areas of the codebase it shouldn't have touched.
Every response the AI sends my way is a claim on my scarce attention layered on top of a tendency toward overwork that was already there, at least according to my therapist. I work from when I wake up until my body stops cooperating. So if someone already wired for overwork can find AI-native work destabilizing, I doubt the problem can be simply attributed to poor discipline. This is the lure of this AI era. Doing more with less now feels morally required, and going to sleep without having hours of work queued up feels like wasteful non-spending. And when it seems like others are living a life of 24/7 agentic building, not following suit feels like failure.
Prior to agentic AI, delegating to teammates came with built-in latency that let me recover. But agentic work keeps opening these little multiverses, each plausible enough to require attention, before my mind has cooled down from the last session.
My bottleneck has shifted from limited time to context fragmentation.
I use AI-assisted development to build Perzimo. AI touches every part of the business. One minute I am using it for writing code or creating design assets; the next I am using it on translations, marketing, or in fundraising. AI has enabled me to build things that would have required several people at my previous startup, but there are still challenges that continue to deep-fry my brain.
AI has increased the volume of work that is falling on me. I used to stress over the accumulation of tasks in the backlog, wishing I could increase the throughput to move more out of it. Now, I stress over the mountain of partial work that gets delegated back to me at the last-mile mark. Every response from the agents arrives to me marked as done, but it rarely is, and I have to spend hours polishing what the agent finished in a few minutes.
AI has the compressed the time between context switches to virtually nothing. I created a skill so I can paste a Sentry URL and get a complete report on an issue that details its root cause, when it started happening, why, how to solve it, along with a TL;DR and ELI5 sections. I'd create 5 Codex sessions with 1 ticket each and a few minutes later, I have five solutions but none were one-shotted. It then falls on me to nudge each agent in the correct direction. It was only a few minutes ago that I started these sessions with a fresh state of mind, and fairly quickly I start drowning in an agentic soup of micro decisions. And if someone were to interrupt me, or if I step away for just ten minutes to grab a snack while a swarm of agents are each doing their own thing, I can come back with no idea what I was doing or what I expected from each of them.
And AI has widened the distance between those context switches. Product to finance. Finance to marketing. Marketing to fundraising. Fundraising to ops. Each jump asks for a different person inside my head, and each person needs a different definition of "good enough." I used to do these domain switches hours apart, now it’s common to do them minutes apart.
That is how the agents show kindness. They rarely say no and never make you feel guilty for asking them to redo a task because your definition was wrong or incomplete. If you throw an expletive at them in anger, they absorb it as if you did not say it. This sounds like a dream until you realize it leads to... so damn much!
A lot of my first responses made things worse. I ran more agents; trusted AI output too quickly; and treated every response as urgent because it looked finished enough. Mostly, I tried to out-discipline a system designed to produce infinite next steps.
As I adjusted, I found some things partially worked: fewer concurrent agents; tighter task scopes; batching reviews; strong evals; fresh sessions; no task starts without a definition of done; and preventing agents from expanding scope just because they find something interesting. Strict AGENTS.md rules helped the most. But the AI still skips rules, under-applies them, short-circuits root-cause analysis, and ends prematurely despite being begged not to.
I don't want to go back to the old world. But sometimes, when I'm overstressed and bleary-eyed, I quietly hope Anthropic and OpenAI hit a few scaling bottlenecks, just so I can catch my breath.