Agents took the bricklaying, and the bricklaying was where my head rested between decisions. Batching the thinking gives some of that rest back.
Last week I shipped six tickets in an afternoon and closed the laptop more tired than after a week of building by hand. Nothing went wrong. Every PR was green. I had spent the whole day deciding things.
Before agents, my day had a shape. I'd plan for an hour, then lay bricks for two: the styled component, the hover state, the fourth breakpoint. Building was familiar enough that my hands knew the way, and that was the rest. Nobody called it rest. It was the part between the thinking.
Agents took the bricks. What's left is planning, setting guardrails, reading diffs, checking screenshots and sorting the bug reports that pile up while they work. All of it sits above the dashed line. My output went up. My head stayed the same size.
The cost hides well, because the day looks great on paper. More shipped than ever, and at seven in the evening I can't decide what to cook.
My first instinct was to stay close to the agents. Watch every run, catch the wrong turn at minute two instead of minute forty. It feels responsible. It's how I'd sit next to a junior designer in their first week.
It also means I never leave the top of the chart. Every glance at a running agent is a small review. I'd notice something, start forming an opinion, and by the time the next ticket landed I'd be holding three half-decisions. Lunch went too, phone out, checking a run from the table. I even set agents up to keep going while I commute and eat, so no time is wasted. No break either.
What works for me now is a queue. Through the day I collect ideas: things I notice while using the product, screenshots, bug reports, a better setup someone showed me. Then I do the high-level thinking in one sitting. Each idea becomes a ticket with a clear spec, and each ticket gets its own worktree, so nothing reaches main before its checks pass. I send the batch and step away.
A batch takes 30 minutes to an hour. That gap is the new bricklaying, so it has to be useful and familiar enough to run without much thought. For me that's a small code experiment that keeps my hands sharp, a walk through the product looking for inconsistencies to feed the next queue, or ten minutes outside in the sun. Vibey music helps. Taking it one step at a time helps more.
Then I come back to a full batch, review it in one go and plan the next one. The day gets its troughs back. They're shorter than the old build stretches, and I'll take them.
The hard part is leaving. I'll admit the wait makes me anxious. An hour is a long time to not know whether the thing worked, and the pull to check is constant. So here's the same hour twice, sped up. In one I leave the batch alone. In the other I check on it three times.
A check is never just a look. I see a half-finished diff, start judging it and carry that judgment back into the break, so the demo charges ten minutes of rest for each one. Leave the agents alone and the whole hour is rest. Check on them three times and it drops to 30 minutes, in scraps of 20 or less. The tickets ship at the same minute either way.
Some work shouldn't be batched. When I'm tuning motion, a spring that feels wrong at 0.3 and right at 0.25, I want to be in the loop on every frame, and an agent gone for an hour on that is an hour wasted. Same with a bug I can't explain yet. The agent needs me steering, one hypothesis at a time.
And a batch is only as good as the sitting that planned it. A fuzzy spec comes back as six confident wrong answers. If I can't write the ticket clearly, it isn't ready to queue.
I like the speed, for the record. I'd take faster batches tomorrow. What I'm after is a cadence that keeps the ideas and the output as good as the speed allows.
Maybe this is easier as a design engineer. The gaps can still hold something I make with my hands, and that part is fun.
Bricklaying was where my taste rested between decisions. Agents never need that rest. That's why I schedule mine.