I triggered an AI review on a PR and went off to do something else. When I came back to my "burn down" list, I'd forgotten which ones I'd started reviews on and which ones I hadn't. So I jumped in, re-read it, and saw the review was still in progress. I did this with three PRs in rotation.
If I were manufacturing a physical product, I'd have just picked up the same part twice[1].
The minutes are the cheap part. The cost is reloading the whole change into my head: what it was for, what finished looked like, what I was worried about.
PRs are inventory, and the lean software crowd has said it for twenty years: unshipped code is stock sitting on a shelf[2]. But with AI the shelf fills fast, and all of it is waiting on me.
As my co-founder Didier used to point out many years back, Toyota worked this out on the factory floor decades ago. The Toyota Production System treats inventory as waste rather than an asset. Kaizen is small, continuous improvement made by the people doing the work. And anyone on the line can stop it when something goes wrong, so the problem gets fixed at the source instead of being passed downstream[3].
For me that's three habits.
(1) Limit Work-in-Progress (WIP). I'm trying, and it's hard. Agents make starting so cheap that inventory piles up before you notice. Epics help, but every open PR is still a context I'll eventually have to reload[4]. In software we adopted the Kanban board, but in my experience the most important lesson was lost: Kanban was meant to physically cap WIP.
(2) Never pick something up twice. If I touch it twice, something upstream needs fixing. Automate it, hand it to a bot, or find a better way so it doesn't come back. This is the double handling problem, but treated as a rule rather than an aspiration.
(3) Stop the line and fix the tool. When a PR goes wrong, don't just fix the PR. Fix whatever let it through: the prompt, the skill, the lint rule, the harness. That's what the Axe Branch is for, and it's why I force myself to spend half of every day on the harness and half delivering.
AI does make kaizen easier. Fixing the process used to mean a ticket, sprint planning, standups, reviews. Now it can be a prompt, and a mistake-proofing lint rule takes minutes. A lot of software can simply be done now, with no scoping and no ticket. That bug that languished on the backlog for years, picked up and put down a dozen times, is stale inventory too. Let the system clear it, and you get room for the work that needs actual thought.
But at first it makes things way, way worse. AI multiplies inventory before it improves anything. More PRs, more re-opens, more half-finished threads. The bottleneck moves to you. Add agents without stopping the line and you just get a bigger pile of stock.