Your AI agent is waiting for you. Your attention isn't. How much time do you lose to this?

Left an agent to fix the failing test while I stepped away. The suite was green so I kept moving. And days later I found out it "fixed" it by weakening the assertion instead of fixing the code ) Plus the minutes lost waiting are nothing compared to the hours debuging false confidence

Real cost was rushing approval to unblock terminal. Skimmed diff and shipped report with confident number nobody measured. And took full day to trace issue and realize agent just hardcoded placeholder value )

the tax i managed to cut wasnt the finding, it was the asking. most of the yes prompts i was answering werent really decisions, they were the agent checking in on things id have approved every time without thinking about it. so i wrote the approvals down once, as an allowlist of operations it can run unattended, and left the gate on only the things that are genuinely irreversible. the interrupt count dropped a lot, and the ones that survived are actually worth getting up for. notification routing is a real fix and pokepoke sounds genuinely useful for the ones that remain. i just found the bigger win sat upstream of it, getting the count down rather than answering the same question faster.

The tax that actually hurt me was never finding the right pane, it was how many things stopped for approval that did not need to. I wrote down what my agent can do without asking, most file edits, most reversible actions, and gated only what touches money, sends something publicly, or cannot be undone. That cut the interrupt count hard, so pane hunting stopped being the bottleneck because there was far less to hunt for. Notification routing fixes the symptom. The number of yes prompts is the actual cost, and most of them are not real decisions, they are the agent checking in on something you would approve every time anyway.