--- schema: 1 type: video title: >- Building uReview, Uber’s Multi-Agent Code Review Engine — Will Bond & Ameya Ketkar, Uber description: >- In 2024 an Uber engineer waited about three hours for a first review on a pull request. In 2026 that wait is nine hours. Volume and size both grew, and code review became the bottleneck for thousands of engineers spread across hundreds of teams, twelve sites and six language specific monorepos. Will Bond and Ameya Ketkar walk through uReview, the system Uber built rather than bought, partly because most vendors do not support Phabricator and partly because they wanted agents in the inner loop reviewed against exactly the same rules as humans. The instructive half is what they had to measure before it worked. Early observability was cost, an NPS survey and a Google form, and the quality to cost ratio landed all over the chart. Tracking reply sentiment, whether a comment actually got addressed, and the agent's own trajectory is what let them tune it, because a model never signals that it is wrong and will assert a bad review with full confidence. Teams write their own reviewers, and Ketkar is blunt that authoring a skill was the easy part while running skills at scale cheaply was not. It now posts about 25,000 comments a week, roughly 67% get addressed, and cost fell 60% against their naive first build. Speaker info: Will Bond: - https://x.com/wbond - http://linkedin.com/in/wbond Ameya Ketkar: - https://www.linkedin.com/in/ameya-ketkar - https://scholar.google.com/citations?user=6JO46GMAAAAJ&hl=en Timestamps: 0:00 - Three hours to review in 2024, nine in 2026 1:53 - Why Uber built this instead of buying it 3:42 - The architecture, and deduplicating comment volume 4:37 - Humble beginnings, and cost as the only metric 5:33 - Sentiment, addressal rate, agent trajectory 6:32 - The model never knows that it is wrong 7:26 - Letting hundreds of teams customize reviews 10:10 - Results: 25,000 comments a week 11:08 - Inner loop versus outer loop 13:57 - Expanding the outer loop rather than killing it --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/EL123UNokkI-youtube-6b5fc2bd09d9cd27bedf Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: 'https://liquidgold.cc/watch/EL123UNokkI' markdown: 'https://liquidgold.cc/watch/EL123UNokkI.md' llms: 'https://liquidgold.cc/watch/EL123UNokkI/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T04:28:56.315Z' language: null children: [] videoId: EL123UNokkI durationSeconds: 907 publishedAt: '2026-08-28T14:30:12.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 24 endSeconds: 114 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=24' heldReason: rights - ordinal: 1 text: null startSeconds: 24 endSeconds: 114 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=24' heldReason: rights - ordinal: 2 text: null startSeconds: 24 endSeconds: 114 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=24' heldReason: rights - ordinal: 3 text: null startSeconds: 114 endSeconds: 126 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=114' heldReason: rights - ordinal: 4 text: null startSeconds: 126 endSeconds: 228 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=126' heldReason: rights - ordinal: 5 text: null startSeconds: 126 endSeconds: 228 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=126' heldReason: rights - ordinal: 6 text: null startSeconds: 228 endSeconds: 240 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=228' heldReason: rights - ordinal: 7 text: null startSeconds: 240 endSeconds: 286 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=240' heldReason: rights - ordinal: 8 text: null startSeconds: 286 endSeconds: 298 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=286' heldReason: rights - ordinal: 9 text: null startSeconds: 298 endSeconds: 342 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=298' heldReason: rights - ordinal: 10 text: null startSeconds: 353 endSeconds: 399 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=353' heldReason: rights - ordinal: 11 text: null startSeconds: 353 endSeconds: 399 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=353' heldReason: rights - ordinal: 12 text: null startSeconds: 399 endSeconds: 410 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=399' heldReason: rights - ordinal: 13 text: null startSeconds: 410 endSeconds: 455 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=410' heldReason: rights - ordinal: 14 text: null startSeconds: 455 endSeconds: 465 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=455' heldReason: rights - ordinal: 15 text: null startSeconds: 465 endSeconds: 621 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=465' heldReason: rights - ordinal: 16 text: null startSeconds: 465 endSeconds: 621 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=465' heldReason: rights - ordinal: 17 text: null startSeconds: 465 endSeconds: 621 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=465' heldReason: rights - ordinal: 18 text: null startSeconds: 465 endSeconds: 621 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=465' heldReason: rights - ordinal: 19 text: null startSeconds: 465 endSeconds: 621 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=465' heldReason: rights - ordinal: 20 text: null startSeconds: 621 endSeconds: 633 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=621' heldReason: rights - ordinal: 21 text: null startSeconds: 633 endSeconds: 679 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=633' heldReason: rights - ordinal: 22 text: null startSeconds: 633 endSeconds: 679 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=633' heldReason: rights - ordinal: 23 text: null startSeconds: 691 endSeconds: 847 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=691' heldReason: rights - ordinal: 24 text: null startSeconds: 691 endSeconds: 847 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=691' heldReason: rights - ordinal: 25 text: null startSeconds: 691 endSeconds: 847 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=691' heldReason: rights - ordinal: 26 text: null startSeconds: 847 endSeconds: 857 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=847' heldReason: rights - ordinal: 27 text: null startSeconds: 857 endSeconds: 907 watch: 'https://liquidgold.cc/watch/EL123UNokkI?t=857' heldReason: rights --- # Building uReview, Uber’s Multi-Agent Code Review Engine — Will Bond & Ameya Ketkar, Uber In 2024 an Uber engineer waited about three hours for a first review on a pull request. In 2026 that wait is nine hours. Volume and size both grew, and code review became the bottleneck for thousands of engineers spread across hundreds of teams, twelve sites and six language specific monorepos. Will Bond and Ameya Ketkar walk through uReview, the system Uber built rather than bought, partly because most vendors do not support Phabricator and partly because they wanted agents in the inner loop reviewed against exactly the same rules as humans. The instructive half is what they had to measure before it worked. Early observability was cost, an NPS survey and a Google form, and the quality to cost ratio landed all over the chart. Tracking reply sentiment, whether a comment actually got addressed, and the agent's own trajectory is what let them tune it, because a model never signals that it is wrong and will assert a bad review with full confidence. Teams write their own reviewers, and Ketkar is blunt that authoring a skill was the easy part while running skills at scale cheaply was not. It now posts about 25,000 comments a week, roughly 67% get addressed, and cost fell 60% against their naive first build. Speaker info: Will Bond: - https://x.com/wbond - http://linkedin.com/in/wbond Ameya Ketkar: - https://www.linkedin.com/in/ameya-ketkar - https://scholar.google.com/citations?user=6JO46GMAAAAJ\&hl=en Timestamps: 0:00 - Three hours to review in 2024, nine in 2026 1:53 - Why Uber built this instead of buying it 3:42 - The architecture, and deduplicating comment volume 4:37 - Humble beginnings, and cost as the only metric 5:33 - Sentiment, addressal rate, agent trajectory 6:32 - The model never knows that it is wrong 7:26 - Letting hundreds of teams customize reviews 10:10 - Results: 25,000 comments a week 11:08 - Inner loop versus outer loop 13:57 - Expanding the outer loop rather than killing it --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/EL123UNokkI-youtube-6b5fc2bd09d9cd27bedf Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/watch/EL123UNokkI · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.