Seven hundred firms turned on coding agents and watched their engineering dashboards climb. That expectation is now measured rather than guessed. Chen and Stratton’s working paper uses Jellyfish data, 718 firms and 300 million work events from January 2021 to March 2026; agent adoption starts near zero in October 2024 and passes 95 percent of the firms by January 2026, so the before-and-after is the firms’ own history. Agents lift lines of code 30 percent, commits 20 percent, pull requests 23 percent. Assistants, adopted earlier, move the same rows much less: +12, +9, +5, with only the commit effect significant. Every number in this note is theirs.

The rows that track shipped work stayed put. Resolved Jira issues and epics, the things that close when code reaches production, show small, statistically insignificant gains; the estimates rule out output growth above 12 percent of baseline. The obvious objection is that agents started taking on bigger issues, so the authors predicted issue length with a model and found no change. Headcount did not move either: a precise zero overall, ruling out declines beyond 2.9 percent.

The extra code queued. Days between pull-request submission and merge rose 3.45 on a 7.03-day baseline, 49 percent longer, and that figure is calendar time, so waiting sits in it alongside reviewer minutes. The share of PRs drawing a formal request-for-changes went from 0.13 to 0.25, nearly double. Comments per PR rose 0.58 on 1.66, and the share of workers doing reviews rose 14 percent. Writing was never most of the job; the paper’s survey puts code-writing at 30 percent of an engineer’s week, its imputation at 42.

Then the firms did the obvious next thing and handed review to AI as well. By March 2026 nearly 80 percent of them had AI somewhere in the review process, yet only 23.3 percent of review comments were AI-generated, and only 10.8 percent of pull requests got so much as one AI comment. The longer reviews, the doubled change requests, and the shift of people into reviewing all persisted through that adoption. The humans kept the load. I wrote up Linear’s engineering lead calling bug-finding and logic checks “now mostly covered by agents”; at the firm level, the review layer around those checks got heavier and more human, not lighter.

What I keep from it is which rows moved. Lines, commits, and pull requests are the default productivity rows of an engineering dashboard, and they are exactly the rows agents inflate by construction. The rows that still tell the truth are in the same tool: resolved issues for output, changes-requested rate and days from submission to merge for the review load. This paper is built on the analytics platform firms already use to watch themselves, so no research team is needed to run the same check against your own history; the honest rows were already in the dashboard. It is a job-market draft, current version August 4, and I have not re-run its regressions. What its numbers add up to: same firms, same headcount, 23 percent more pull requests, each taking three and a half days longer to land.