524 real PRs show what is dragging review down is not AI, it is a complexity bill nobody has put a number on
ReviewDebt: a practical framework for scoring every pull request — Sachin Gupta, Ebay · Sachin Gupta
25 min total·Actually worth watching closely: ~7 min·3 must-watch clips
- 0:00 – 3:35Skim
The gap nobody is measuring
The speaker opens flat out: coding agents now produce code faster than humans can responsibly review it, and nobody is measuring that gap. Three sets of industry numbers follow to pin the problem down.
In the same year, commits climbed 25% while comments on commits dropped 27%; median PR review time is up 441.5% and 31% more PRs are merged with no review at all — production volume and review attention are moving in opposite directions.
Data-dense, but all of it is cited report figures shown as text and charts; skim for the order of magnitude rather than following every sentence.▶ Jump to 0:00Speaker · Sachin Gupta - 3:35 – 8:00Listen
Why it is debt, not technical debt
The definition of review debt: the accumulating gap between the code your agent has produced and the code humans have actually reviewed, trusted and understood. The speaker stresses that it compounds, that the interest is paid in human attention, and lists five hidden costs.
The nastiest of the five is test theater — agent-generated tests assert what the code did rather than what it should do, locking bugs in as settled fact.
Conceptual reasoning and analogy throughout, with the slide parked on a single definition page — you lose nothing listening while walking or commuting.▶ Jump to 3:35Speaker · Sachin Gupta - 8:00 – 13:24Skim
How the ten deterministic checks fit together
The scoring framework unpacked: five signal families and ten checks computable straight from the PR, covering diff size and coupling, test evidence gap, ownership spread, agent authorship indicators, and evidence and rationale gaps, combined into a 0-100 score with four disposition bands.
The framework deliberately keeps an LLM out of the judge seat — only metrics that can be deterministically recomputed, so the score can be challenged, defended, and written into process.
Around 10:52 there is a side-by-side worth pausing on: the high-gap PR's body is 18 characters with a commit message reading 'updates', next to a low-gap PR that spells out symptom, diagnosis, change and a benchmark link. Getting that pair is worth more than hearing all ten check names.▶ Jump to 8:00Speaker · Sachin Gupta - 13:24 – 16:52Watch
Three real PRs through the scorer
Three real PRs from public repos go into the scorer: a well-shaped one scores 0, a high-debt one scores 60, and one written by an agent but shipped with tests and a called-out risky path scores just 7.
What drives a high score is never who wrote it, but size, coupling and missing evidence; the agent only creates the conditions under which the debt piles up fast.
Everything here lives on screen — the individual report lines, how many points each check cost, the estimated review minutes. Audio alone loses all the force of the argument, so watch it.▶ Jump to 13:24Speaker · Sachin Gupta - 16:52 – 20:26Watch
The real bill from 524 PRs
The scan widens to three public repos, 90 days and 524 PRs, showing the accumulation curve of review hours and what the PRs that reach the high burden band actually look like.
Only 4 of the 524 PRs landed in the top two bands, and all four were structural changes such as large migrations; yet 228 senior reviewer hours have already accumulated, with the most extreme single PR estimated at 84 hours.
The accumulation curve and the two-repo comparison only become intuition once you see them — read aloud it is just '186 hours and 43 hours', but only the plot shows which week they started to diverge.▶ Jump to 16:52Speaker · Sachin Gupta - 20:26 – 23:40Listen
The prescription and the five steps to adopt it
The team rules that need no new tooling (one logical change per PR; tests ship with the change and a human confirms they assert intended behavior; cross-cutting work split along team boundaries; the why written by the human author), followed by a five-step adoption path that starts with measurement.
Score last week's 20 PRs first to see whether the framework holds — don't build tooling out of the gate; surface the score as a comment without blocking the merge, and watch the weekly slope of the debt rather than its absolute level.
A spoken checklist where the items matter more than the presentation — listening and taking notes is enough, no need to watch the screen.▶ Jump to 20:26Speaker · Sachin Gupta - 23:40 – 24:57Listen
Three things to take away
A one-minute close: three things you can do tomorrow, a list of anti-patterns to avoid, and a hand-off to the governance questions 2027 will not let anyone dodge — trust, accountability and audit trails.
Turning the code quality conversation from 'this PR feels risky to me' into a number you can actually argue over is what the whole framework is really buying you.
A purely spoken summary with no new visuals; the last minute is all you need.▶ Jump to 23:40Speaker · Sachin Gupta