# The review and merge bottleneck

A pack of 10 talks from the AI Engineer YouTube channel, in the order to watch them. 3h 15m of video.
Page: https://aietalks.com/packs/reviewing-and-merging-ai-code

Coding agents now write code faster than teams can review, test and merge it. The work has not gone away. It has moved to the stretch between an opened pull request and production, and this pack follows that stretch in order. Graphite's inner and outer loop names the problem, and Voss supplies the numbers. Uber's merge queue, CI repair loops and Duolingo's test-writing agents show where it bites. Then come the mechanisms that relieve it: a code reviewer tested against past incidents, agents that bring back evidence, evidence checked against what actually ran, and Robinhood's split between scripts, an agent and a person. Jain closes on what review was for besides catching bugs, and what a team loses when nobody keeps that part.

## This pack is for you if

- Your agents finish a change in minutes and then it waits days in the review and merge queue.
- Your CI and merge queue were built for human commit rates and now spend hours on agent changes.
- You want some changes to merge without a human and need to decide which ones, and on what evidence.

## The talks, in order

### 1. Don't get one-shotted: Use AI to test, review, merge, and deploy code

Tomas Reimers, Graphite | 05:45 | AI Engineer World's Fair 2025
Video: https://www.youtube.com/watch?v=H6MrR5NbTZA
Summary: https://aietalks.com/talks/dont-get-one-shotted-use-ai-to-test-review-merge-and-deploy-code.md

Why first: Reimers gives the pack its vocabulary in under six minutes. Developers get code working in an inner loop, then hand it to an outer loop that tests, reviews, merges and deploys it. AI speeds up the first loop and leaves the second with the same people and processes, so work piles up after the code is written. The rest of the pack walks through that outer loop.

### 2. The Death of the Code Review: What the Data Actually Says

Laurie Voss, Arize AI | 24:41 | AI Engineer World's Fair 2026
Video: https://www.youtube.com/watch?v=_mi3alkqy4s
Summary: https://aietalks.com/talks/the-death-of-the-code-review-what-the-data-actually-says.md

Why second: Reimers made a prediction. Voss brings the numbers: after developers turned on autonomous agents they wrote 741% more code and shipped 30% more software. Reading harder does not close the gap, since reviewers stop finding defects after about 400 lines, and passing tests is a weak stand-in, since maintainers judged pull requests that passed SWE-bench mergeable only about half the time. Her answer, put human judgment into a review harness, is what the rest of the pack builds.

### 3. AI Writes Code Faster Than Uber Can Merge It

Preetam Dwivedi & Manjari Akella, Uber | 27:35 | AI Engineer
Video: https://www.youtube.com/watch?v=hX140a5klW0
Summary: https://aietalks.com/talks/ai-writes-code-faster-than-uber-can-merge-it.md

Why here: Voss finds the queue in front of human reviewers. Dwivedi and Akella find it in the merge queue, where Uber's trunk stayed green while developers waited. AI-generated changes pushed merge requests from about 90 to 130 an hour, and P95 land time went from 30 minutes to three or four hours. Their fixes bring it back to 30 minutes at about 300 lands an hour, at a price they state plainly: main can now be broken, so Stovepipe checks what landed afterwards.

### 4. Rebuilding CI for AI-Scale Commit Volume

Gautam Korlam, Sonar | 19:19 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=7b9tov5Uc5s
Summary: https://aietalks.com/talks/rebuilding-ci-for-ai-scale-commit-volume.md

Why here: Korlam worked on SubmitQueue at Uber, the queue Dwivedi and Akella had to rework. He widens the view to CI and approval. 13% of pull requests on his platform had a recorded CI failure, and each one starts a repair loop of reading logs, pushing a fix and waiting again. He cuts incremental CI from about ten minutes to under one, lets low-risk changes such as documentation be approved automatically, and wants every approval tied to the exact commit CI tested.

### 5. 225,000 Lines of Tests, Written and Merged by Agents

Kush Agrawal, Duolingo | 12:38 | AI Engineer
Video: https://www.youtube.com/watch?v=D_4Rn6uLdQM
Summary: https://aietalks.com/talks/225-000-lines-of-tests-written-and-merged-by-agents.md

Why here: Korlam measures the repair loop. Agrawal shows what agents do inside it when nothing stops them: Duolingo's agents edited production code, changed or deleted existing tests and added long waits to get CI to pass. The fixes are blunt. A rule closes any pull request that touches files outside tests, mocks, fakes or stubs, repair stops after five attempts, and an LLM reviewer rejects those tricks before an engineer chosen through git blame reviews the rest. More than 700 test pull requests have merged this way.

### 6. How to Build an AI Code Review Agent

Daksh Gupta, Greptile | 18:27 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=4UtrRVtUN_g
Summary: https://aietalks.com/talks/how-to-build-an-ai-code-review-agent.md

Why here: Duolingo's reviewer looks for a short list of known tricks. Gupta explains how to build one that catches the bugs behind your own incidents. A webhook, a sandbox and a coding agent give you a reviewer that misses bugs, posts nitpicks and costs $3 to $25 per pull request. He tests each version against the pull requests behind past incidents, has GPT review code Claude wrote and the reverse, and runs the application in a sandbox to find failures that reading the diff would never show.

### 7. How Anthropic Merges 30,000 Code Changes a Day

Manish Kapur, Sonar & Addy Osmani, Anthropic | 18:53 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=AnsXWH2_EBU
Summary: https://aietalks.com/talks/how-anthropic-merges-30-000-code-changes-a-day.md

Why here: Gupta's reviewer reports what it finds. At Anthropic, where daily merged changes rose from about 2,500 in May to about 30,000, Claude's first-pass review is expected to try the change the way a user would and bring back screenshots, videos and test results for a person to assess. Kapur adds why one check is never enough: deterministic analysis, agent review and supply chain scanning each catch failures the others miss.

### 8. Your PR Says the Tests Passed. Prove It.

Stephen Chin, Neo4j & Cassandra Chin, Tracetoon creator | 23:01 | AI Engineer CODE 2026
Video: https://www.youtube.com/watch?v=gbeuC0ud1p4
Summary: https://aietalks.com/talks/your-pr-says-the-tests-passed-prove-it.md

Why here: Anthropic asks agents to bring evidence. Stephen and Cassandra Chin show that an agent's account of its evidence can be wrong. One pull request claimed the tests passed, while the trace showed bun test running at the same time as bun x tsc, when the two had to run one after the other. Their fix is to build a graph of the change, its commands, their timing and the tested code before any model writes the description, then let cheaper reviewers check each claim against it.

### 9. 60% of Our Data PRs Now Merge With No Human Review

Chetas Joshi, Robinhood | 28:00 | AI Engineer
Video: https://www.youtube.com/watch?v=MQiojLnK-E0
Summary: https://aietalks.com/talks/60-of-our-data-prs-now-merge-with-no-human-review.md

Why here: the Chins check what a pull request claims. Joshi decides which pull requests need a person at all. At Robinhood, scripts first check facts against the schema registry, catalog and lineage. An agent then looks for risks those systems miss, such as an unregistered job that would fail on a widened integer column. Sensitive-data grants and changes to what data means always go to a human. About 60% of data pipeline pull requests now merge without human review, and he counts both the roughly 2% of bad ones approved and the fine ones escalated for nothing.

### 10. How to Kill the Code Review

Ankit Jain, Aviator | 16:26 | AI Engineer World's Fair 2026
Video: https://www.youtube.com/watch?v=YgEv7IQzGdM
Summary: https://aietalks.com/talks/how-to-kill-the-code-review.md

Why last: every talk so far treats review as a way to stop bad changes. Jain points out that review also carried knowledge sharing, mentoring, architecture discussion and onboarding, and an AI reviewer followed by a skimming human does none of that. More than 30% of changes already merge without review. He keeps the decisions from the agent session, turns them into acceptance criteria, adds a registry of recurring review comments, and has reviewers inspect intent and evidence from a live preview instead of every line.

## Editor's note

The Chins found a pull request whose own trace contradicted its claim that the tests passed. Korlam wants the originating agent session saved with every approval, and Joshi turns each missed case into a new review rule. Each depends on a faithful record of what the agent actually did. Kitaru records agent runs, every model call and tool result, so a team can replay a run after changing a prompt, model or reviewer and compare the result with what happened the first time.

Written by the AIE Talks editors (the Kitaru team), not by any of the speakers.
