Volume vs. Value

Agents can open ten times the pull requests in a tenth of the time — and that is exactly why your review queue, not your IDE, is now the thing that decides how fast you ship.

Why this, for you: you turned on a coding agent and your PR count jumped. It feels like velocity. But the PRs pile up unmerged, your reviewers are drowning, and the backlog of hard work looks untouched. This lesson names what actually moved — and what didn't — so you stop measuring the wrong thing.

The reflex is to count pull requests opened. That number went up, so productivity went up — right? The large-scale data says the bottleneck just relocated. Generation got cheap; judging what to merge got expensive, and it is now the binding constraint.

1 The paradox: faster output, lower acceptance

The AIDev dataset gives the first large-scale picture of agent-authored PRs in the wild — 456,535 pull requests from five autonomous coding agents (OpenAI Codex, Devin, GitHub Copilot, Cursor, Claude Code) across 61,453 repositories. Two numbers move in opposite directions:

Metric	Human	Codex	Copilot
Median close time	3.9 h	0.3 h	17.2 h
Acceptance rate	~77%	64%	35%

Codex PRs close ten times faster than human PRs. But the best-performing agent still trails humans by 13 percentage points on acceptance; Devin lands at 49% and Copilot at 35%. Speed does not buy back the shortfall in quality and relevance.

One developer submitted 164 Codex-assisted PRs in three days — nearly matching the 176 human-authored PRs that project produced over 3.5 years. The volume is real. Whether it delivers proportional engineering value is the open question this whole course exists to answer.

2 Volume overstates the work that got done

Raw PR count is a misleading proxy for two reasons. First, agent PRs skew structurally simpler: only 9.1% of Codex-assisted PRs changed cyclomatic complexity, versus 23.3% for human PRs. Agents gravitate to routine fixes, documentation, and scaffolding — not architectural work. Many tasks really are simple, but the complexity gap means the headcount-style PR tally overstates how hard the completed work was.

Second, more PRs is not free throughput. Acceptance lags because agents lack the ambient project context that shapes which change is worth making right now — roadmap direction, frozen subsystems, unwritten priorities. So they optimize for local correctness over relevance to the broader work queue.

Merge is not the finish line

Even merged agent PRs over-report value: a study of 1,210 merged agent-generated bug-fix PRs found that merge success does not reliably reflect post-merge code quality — code smells at critical and major severities dominate the introduced defects. And higher volume raises a separate tax: a 142K-PR analysis reports a 27.67% merge-conflict rate. Throughput gains erode once rebase costs are counted.

There is a bright spot, and it is consistent: documentation. Codex (88.6%) and Claude Code (85.7%) beat the human documentation acceptance rate (76.5%). Natural-language generation is a core model strength, and doc PRs cash in on it.

3 The bottleneck moved to review

If you can generate ten times the PRs but accept fewer of them, the scarce resource is no longer authorship — it is the reviewer-hours that decide what merges. The signature is already visible in the data: bots make up 20% of reviewers on agent PRs, double the 10% on human PRs. Agent code increasingly passes through automated review before — or instead of — a human.

As volume rises, the value shifts from writing code to judging it. Deciding what to merge now outweighs producing the change — review becomes the highest-leverage engineering skill on the team.

But you cannot simply hand review to a bot to keep pace. When code-review agents (CRAs) review without a human, the merge rate collapses:

Reviewer composition	Merge rate
Human-only	68.37%
Human-dominated mixed	67.99%
CRA-only	45.20%

That is a 23-point gap across 3,109 PRs — and it contradicts industry claims that CRAs can handle 80% of PRs unaided. The cause is low signal ratio: 60.2% of closed CRA-only PRs got fewer than 30% actionable comments, and 12 of 13 CRAs studied averaged below 60% signal. When most comments are noise, developers stop acting on them, PRs stall, and they are abandoned. (Why throughput beats coverage is Lesson 2.)

# The metric swap that changes how you steer - track: pull requests opened # up 10x — and misleading + track: merge rate (accepted / opened) + track: reviewer-hours per merged PR + track: signal ratio of automated comments

↪ Your win: count value, not volume

Volume is not value — agents open ~10x the PRs but trail humans 13–42 points on acceptance.
Merge rate, not PR count, is the productivity metric for agent-assisted work.
Volume overstates difficulty — agent PRs change cyclomatic complexity 2.5x less often (9.1% vs 23.3%).
The bottleneck is review — generation got cheap; judging what to merge is now the binding constraint.
Don't drop the human — CRA-only review merges at 45% vs 68% for human-involved review.

Retrieval practice — recall, don't peek

Question 1In the AIDev data, agents open far more PRs than humans but their acceptance rate is…

Question 2The single metric that best measures agent-assisted productivity is…

Question 3Agent PRs change cyclomatic complexity far less often than human PRs (9.1% vs 23.3%), which means raw volume…

Question 4When code-review agents review without any human involved, the merge rate (45%) versus human-involved review (68%) shows you should…

Question 5With agent output rising, the highest-leverage engineering skill shifts toward…

Ask me anything. Want help picking a merge-rate baseline for your team, or auditing where your review queue actually chokes? Next, in Signal Over Volume: why optimizing review for coverage backfires — and the disproportionate security-review gap hiding in AI PRs.