Throughput went up. My evenings got worse.
I build a product mostly on my own, with agents doing a growing share of the mechanical work. This year my output went up and my days got worse, and it took me a while to see why.
Automation ate the typing. It did not touch the checking, and it multiplied the number of things waiting to be checked.
Then I found telemetry that matches: Faros AI looked at two years of workflow data from 22,000 developers across 4,000+ teams. Task throughput per developer up 33.7%. Median time a pull request sits in review up 441.5%. Daily PR contexts per developer up 67.4%.
Two more numbers from the same report show where the pressure goes when nobody has capacity: PRs merged with no review at all, up 31.3%. Incidents per PR, up 242.7%.
The reason it costs so much feels obvious once you name it. Writing builds the mental model as you go. Review is rebuilding that model backwards from a finished result, without the branches the author threw away. A colleague signals uncertainty — a hedge in the commit message, a comment, a ping. Generated work arrives uniformly confident, so every section gets the same level of attention.
What helps me is unglamorous: smaller units of generated work, fewer parallel threads even at the cost of throughput, tests written before generation, and treating "I don't understand why this works" as a blocking result rather than a style nit.
None of it closes the gap. It moves it a little.
For those of you shipping with agents daily — where does your review time actually go, and what have you changed to make it survivable?
Replies
This matches my last quarter exactly. Solo-building a side project with Claude Code, commit rate roughly doubled and my satisfaction dropped hard.
Where my review time actually goes: ~60% on the "why did it pick this path" question — rebuilding the mental model backwards, exactly like you named it. The other 40% is timezone / empty-state / null-check edge cases the agent glosses over with confidence.
Two changes that moved it a little: blocked 30 min at a fixed daily time just for reviewing yesterday's agent output, and started writing one-line intent comments before generating the code so review has an anchor instead of reverse-engineering the goal.
The 441% PR review number felt inevitable once I saw it. The bottleneck moves from typing to trusting.