---
title: 'In the News: November 20, 2026 (Extra 3)'
description: 'Runway lets agents approve its low-risk pull requests after a trial in which people objected to 15 of 977 AI approvals.'
canonical_url: 'https://darkfactory.dev/news/2026-11-20-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-20-extra-3.md'
collection: news
date_published: '2026-11-20T23:01:00-05:00'
date_modified: '2026-11-20T23:01:00-05:00'
---

# In the News: November 20, 2026 (Extra 3)


Runway now lets an agent approve low-risk pull requests with no person involved, and it published how it decided that was safe. The reviewer approved alongside people first, the company counted how often a person disagreed and how often changes were reverted, and only then let AI approval count on its own.

## 1. Runway lets agents approve its low-risk pull requests after measuring how often people disagreed

**[Building a Software Factory](https://runway.com/news/engineering/building-a-software-factory)** · Alejandro Alvarez, Technical Staff, Runway · Runway Engineering, November 20, 2026

Runway lists seven jobs between a task and production (write, pass CI, review, approve, QA, fix breaks, run the line) and describes handing each one to agents. An internal service runs cloud coding agents that stay with a pull request until it is green. It drove 38% of merged pull requests in October, up from 0.1% in June, and in a sample of 100 of its pull requests merged in the first half of November, 78 had no commits from a person. An AI reviewer with the same context as the coding agents now reviews every pull request, about 3,000 a week across six repositories.

Approval has the most careful evidence. A separate agent rates each pull request low, medium or high risk, "rounding up when unsure." Low-risk changes can merge on AI approval, medium-risk ones can where a repository allows it, and high-risk ones go to a person. Code owners still sign off on database migrations, CI and deploy configuration, secrets and authentication. Across September and October, while the reviewer approved alongside people, a person objected to 15 of the 977 pull requests it approved, about 1.5%. The rate stayed about the same after mid-October, when AI approval alone began to count. Runway reports that about one in a thousand pull requests merged on AI approval alone from October to November was reverted, against about one in two hundred approved by people, and says AI approvals held comparable to human ones when it compared within risk levels. The median wait for a first approval fell from 2.3 hours in September to about five minutes in mid-November.

The later stations were handed over the same way. A QA loop in which agents drive browsers through the live product found 86 confirmed bugs in its first four days and fixed 74 of them in that window. Since early November, every merged change has to pass a verification run on its own preview environment first. A coordinator that reads the issue tracker, hands out work and orders merges ran read-only at first, and a person judged every action it proposed in short test runs. In its first two weeks live it dispatched over 14,000 agent tasks, which led to about 1,900 pull requests. The post lists five safety rails, each tied to a failure Runway hit. One is re-checking gates at the exact commit being merged, added after a pull request was merged by override when its checks had refused it. The rule Runway says it applied throughout: "We decide how each station will be judged before we look at its results."

Every figure comes from Runway's own systems, in a post that ends with a hiring call. We have not seen any of them checked from outside the company.

**Why it matters:** Runway earned agent approval with a measurement. It ran the reviewer in parallel with people, counted objections and reverts, and kept owner sign-off where a mistake costs the most. A team planning to remove a human approval gate can copy that order: decide the measure, run the agent in shadow against people, then hand over only the risk tiers the numbers support.
