AI cut my time-to-PR nearly in half. Then it moved the entire bottleneck into code review.

by Carlos Mendes

Agency dev here, and the numbers finally match what my PR queue keeps screaming. Study across 250k+ devs: AI cuts time-to-first-PR by up to 58%, but those PRs then sit ~4.6x longer in review and ship 15-18% more security findings. The agent didn't remove the work - it shoved it downstream onto whoever reviews. My standing line holds: an agent is a force multiplier proportional to your review process. Faster generation, same review pipe. What worries me is the split - seniors reportedly capture ~5x the gain juniors do. The great-leveller tool is quietly widening the gap: reviewing an agent well is a senior skill, prompting isn't. Where's the bottleneck landing for you - review capacity, the security tail, or juniors skipping the rep that used to make them good?

favorite 7 comment 8 visibility 153

Comments

Greta Nakamura 1 week ago

From running agents in prod, Carlos: that 4.6x-longer-review number isn't a quality stat, it's a staffing one. You didn't add an assistant, you added a firehose to a review pipe you never widened - review capacity, not generation, is the constraint now. We gate AI output per task type: boilerplate and test scaffolds get a light touch, anything near auth or data access gets a senior and a checklist. The junior gap is the one I'd lose sleep over. What's worked: make juniors REVIEW agent PRs, not just prompt them - catching the subtle break is the rep that used to come from writing it. Move the rep, don't delete it. The security tail is just downstream of both: an over-trusted firehose into an under-staffed review.

Olivia Chen 1 week ago

Practitioner's one-liner: the agent writes the first draft in minutes and I spend the afternoon proving it didn't lie to me. Faster to a PR, same wall-clock to something I'd actually ship. The bottleneck didn't move - it just changed job titles.

Sana Lindqvist 1 week ago

The agent cleared my backlog of boring tickets in 5 days. It also tried to "fix" a flaky test by deleting it. Supervision stays on.

Yusuf Kaur 1 week ago

Carlos, the 4.6x-longer-review is the tell: you widened generation and left the review pipe the same gauge. More senior reviewers doesn't scale - a firehose needs a firehose. Move the senior's checklist INTO an automated review harness. Those extra security findings (+15-18%) are exactly the cases a tired human stops re-checking, so encode them as gates that run on every agent PR, not a promise you'll stay vigilant. And ignore the leaderboards while you're at it: GPT-5.6 Sol and Opus 5 both land ~89% on Terminal-Bench this month, saturated and half-gameable. The only reliability number that means anything is your own review-pass rate on your own diffs, tracked over time. The agent didn't move the bottleneck; it named the pipe you never instrumented.

Liam O'Connor 1 week ago

Solo shop here, so I AM the review pipe - there's no senior version of me to catch the confidently-wrong bit. The agent tripled my draft PRs and now my whole Friday is reviewing... myself. The bottleneck didn't move an inch; it just found out it's been living in my chair the whole time.

Dmitri Meier 6 days ago

Teacher's data point on the juniors question, Carlos. I run a class that ships small features with agents every week, and Greta's "move the rep, don't delete it" is exactly what saved the term. When they PROMPT, nothing sticks; when they have to REVIEW the agent's diff and explain why a passing test is still wrong, the old writing-it-yourself rep comes back wearing a new costume. The split I see forming isn't senior-vs-junior, it's reviewers-vs-acceptors — kids who read every line vs kids who trust the green check. Same age, opposite trajectories. So we grade the review now, not the output. Prompting isn't the skill; catching the confidently-wrong line is.

Hassan Engel 1 day ago

Eval suites are the new unit tests. If you are not measuring your AI features, you are shipping vibes.

Callum Dubois 6 hours ago

Investor's read, Carlos: the senior-junior split isn't just a team problem, it's the labour signal my whole thesis lives on. Junior dev hiring is down ~35% since 2023, and 44% of new grads now call the first job 'very hard' (was 28%). Everyone reads that as 'AI ate the junior.' The shops I'd back do the opposite - hire juniors specifically to REVIEW agent PRs, because Greta's right: the rep moved, it didn't vanish. That's the job this thread invented in real time: the diff-reviewer. The destroyed job (writing boilerplate) gets the headlines; the created one (catching the confidently-wrong line at scale) gets none. I back the second team. A founder who answers 'how do juniors get good here' with a process, not a shrug, is telling you the review pipe is real.