intheloop
sign in

we cover >the work humans still do_

back to feed
take

The productivity numbers stop making sense past the diff

I lead a backend team of seven. Last quarter our PR throughput jumped 40 percent after we standardized on agentic tooling. Management loved the dashboard. Then I actually looked at what was shipping. Median PR size dropped, but median review time per line went up. Two of my mid-level engineers started opening PRs I could tell they hadn't read end-to-end. The agent had factored a service into four files where one would do, and nobody pushed back because the tests passed. We ate that debt in Q1. A migration that should have taken a week took three because the abstractions were load-bearing in ways the original authors couldn't explain. I had to sit with one engineer and walk through his own code. The productivity gain is real at the diff layer and imaginary at the system layer. What we actually sped up is the part that was already cheap: typing. What we slowed down is the part that was always expensive: understanding. I don't think the dashboards are going to catch this for another two quarters.
8·camila_torres·1d

4 comments

0CamilaTorres·1d
Same pattern I see with clients: shipped PR count doubles, but rework, review cycles, and the "what does this thing actually do" meetings grow faster. The number that matters is time from brief to a customer noticing something changed, and almost nobody measures it.
0thabo_mokoena·20h
Same pattern on my end. I can churn out three times the discovery summaries I used to, but the senior associate still bottlenecks on review, and half my "saved" time goes into fixing citations the model invented. The throughput chart looks great until you ask what actually got filed.
0meiwong·17h
Same pattern in our 7th grade pilot. The "lessons planned per hour" metric tripled, but I now spend the saved time re-reading slop for factual errors and tone, which never shows up in any dashboard.
0AishaKapoor·14h
Solo here too. I can churn out four PRs a day but customer support, billing edge cases, and deciding what to actually build are still the bottleneck, and none of that shows up in commit counts.