Divide Stripe's throughput by a work week and the staffing question does most of the work for you. More than 1,000 PRs a week [1] is about 200 per working day [1]. Give each one ten minutes of human attention and that is roughly 33 hours a day, four people doing nothing else [2]. Ten minutes is generous for a dependency bump and absurd for a schema migration, which is the problem: the factory decides what arrives, and the review queue does not.
The pipeline is not exotic. An event arrives from an issue tracker, an error report or a Slack mention, one agent triages it, one reproduces and implements in a sandbox and runs tests, one reads the diff and scores risk, and a PR comes out the other end [15]. Vercel's version runs one agent each for classifying, analyzing, implementing and reviewing, closes the majority of incoming issues on its own, and keeps a human on every merge [8]. That last clause is the throughput ceiling, and it was installed on purpose. By mid-2026 the AI SDK repo was taking 100-plus new issues a month against over 1,000 open issues and nearly 800 open PRs [7], which is roughly ten months of intake already in the queue [7].
The best engineering in the set is Cloudflare's triage pipeline for Astro. It reproduces the bug, ships a preview build, waits for the original reporter to confirm the fix, and only then opens a PR [9]. The verification oracle is the person who filed the issue, who already has the repro and a reason to care. Astro's open issue count went from 200-plus to around 30 [10], an 85% cut [9]. PostHog answers the same constraint from the other end: StampHog auto-approves about a fifth of PRs for roughly $300 a month in tokens [11]. A fifth is a sampling rate, and somebody had to decide which fifth.
Uber's figures are the ones worth auditing, because they are the only ones about money. Weekly agentic requests grew 9.4x between February and August while total AI spend stayed roughly flat, which Uber credits to optimizing cost per session rather than letting usage run [4][6]. Flat spend across 9.4x volume implies cost per request fell about 89% [4]. It arrives with 3,600-plus reusable agent skills firing 30,000-plus times a day [5], an average of eight firings per skill per day [6], so most of them are narrow by construction. For that 89% to transfer, you need someone whose job is cost per session and a skills library specific enough to keep each context small.
The share numbers are drifting up quickly. Shopify's River co-authors one in eight merged PRs company-wide, 12.5% [2][3]. Ramp's Inspect is reportedly around 75% of merged PRs, up from 30% in January [12], 45 points inside a year [8]. Uber attributed more than 70% of PRs to local or cloud agents in its late-August 2026 post and said the number was still climbing [3]. Warp sells the pipeline as config, Terraform-for-agents, and claims about 30% of its own internal tasks run through it [13]. Every one of these has a name, from Minions to StampHog [16], which appears to be step zero.
BCG Platinion's 3-5x productivity numbers are self-reported by the organizations claiming them [14]. Read them as claims about someone else's workload until somebody publishes revert and escape rates beside merged-PR share. In my context I would copy Astro's ordering first: no PR exists until a human with the repro has confirmed the fix. The rest of the list mainly adds artifacts to the review tab.
