2026-06-25
By Vadym · Generated with Boba, curated by me
Six weeks ago I wrote that AI brain fry was real and that I had it. I expected writing it down to take some of the charge out of it. It didn't. What it did was make me notice how often I hit 6pm fried and then sat back down at 8pm to start over, because "the agents were still running" and that felt like a reason.
Less diagnosis this time. More what I did about it.
When I wrote the May post, the cutting edge was running a few agents in parallel. That already feels quaint. In June, Anthropic shipped Dynamic Workflows for Claude Code — a mode that plans a task, fans it out across a large number of subagents, verifies, and iterates, with an explicit warning that it burns far more tokens than you're used to. Agent Teams landed too: multiple coordinated instances with one lead. The unit of work is no longer "me and an assistant." It's me supervising a team I don't manage the way I'd manage people.
The research caught up to the feeling. The study I gestured at vaguely in May has a name now: BCG's work, written up in Harvard Business Review in March, found that intensive oversight of AI was the single most mentally taxing way to engage with it — workers doing close monitoring reported 14% more mental effort, 12% more fatigue, and 19% more information overload than people using AI lightly. The ones in that state made major errors 39% more often. Oversight, not output, is where the cost lives. That's the whole thesis of the last post, measured.
The trust numbers stuck with me more. Stack Overflow's 2026 developer survey put AI adoption at 84% and trust in AI accuracy at 33% — only 3% said they highly trust it. We're using it more than ever and believing it less than ever, and the gap between those two lines is exactly the validation tax I keep paying. Google's DORA team even named it that this year: a "verification tax," reviewing generated code as overhead that eats into the productivity it promised.
So: not a personal weakness, not getting better, and now mainstream enough to have line items in the big annual reports.
I listed five practices in May. Six weeks of honest testing:
Capping the agents held. This is the one that actually works. Three concurrent, 90-minute blocks. Past that I'm triaging, not engineering.
Batching reviews held, mostly, except when a release is on fire.
Calibrating trust per task held. It's just a habit now.
"Don't skip the boring human parts" half-held. I skip them exactly when I'm most fried, which is exactly when skipping them costs the most.
"Know when to go offline" did not hold at all. This was the one I admitted I was worst at, and writing that down changed nothing. The problem isn't that I don't know I should stop. It's that there's no moment that tells me to. The agents don't get tired. The queue doesn't empty. Nothing in the workflow has an end-of-day. So I'd drift past it every time.
That's the gap I wanted to close. Not a burnout cure — a stopping signal.
I built a small iOS app called Lagrange Reset. Free, runs entirely on-device, no account. I built it for me first, then figured other people in the same spot might want it.
It's deliberately tiny. At the end of the day you do a two-tap reset: log how you worked (led / collaborated / leaned on the agents) and how you feel (good / ok / fried). That's the whole daily interaction — about ten seconds. The point isn't the data. The point is that it's a deliberate full stop. A line you draw that the workflow won't draw for you.
If you tap fried, it opens a short recovery menu. And this is the part I actually cared about getting right: the options are tuned to this kind of tired, not generic wellness. A three-minute look-away. A no-agent walk — phone stays, no terminal in your pocket. Text an actual human. A wind-down playlist. "A no-agent walk" instead of "go outside," because the thing draining you isn't a lack of steps, it's that you haven't had a thought that wasn't a code review in nine hours.
Then it does one quiet useful thing: it tracks which recoveries actually correlate with your energy the next day. Not what's supposed to work. What works for you. If the walk helps and the playlist doesn't, it'll tell you, and you can stop pretending the playlist helps.
A Lagrange point is the one stable spot between two big gravitational pulls, where the forces balance and an object can hold its position without being dragged into either. The James Webb telescope sits at one. It doesn't fight the pull — it makes small, regular corrections to hold station.
That's the model. The two pulls are the work and you. Most days the work wins by default because it's the one that's always pulling. The app is the small, regular correction — not a dramatic intervention, just enough to keep you from getting dragged all the way in.
It is not a burnout cure. It is not a productivity tool — it will not help you ship more, and if anything it's nudging you to ship less, later in the day. It's a boundary tool. One job: help you stop, and help you recover when you're fried, with suggestions aimed at the specific fatigue that comes from validating machines all day.
In May I ended by saying I was going to close the laptop, and that part was still mine. It turned out I needed help with the closing-the-laptop part. So I built it. It's on the App Store now (Lagrange Reset), and it cost me one small, regular correction.