Working Surfaceissue · 19 Nov
Issue · 5 links
Issue

Thursday
19 November 2026

Five pieces from 17 and 18 November 2026 describe teams replacing human review of every pull request with rules about which changes a person must check, as designers at Salesforce begin opening pull requests of their own. They also show that those rules, and the instructions given to AI agents, need owners and testing of their own.

01

Rootly

When should AI be allowed to approve a pull request?

Quentin Rousseau · 17 Nov

Rootly's chief technology officer found that 90% of recent pull requests were approved without a human comment, and proposes letting an AI reviewer approve only low-risk changes.

Company story · critique & review · decisions · Human approval

02

The Pragmatic Engineer

Software factories and code reviews with Harjot Gill

Gergely Orosz with Harjot Gill · 18 Nov

CodeRabbit's chief executive Harjot Gill says his engineers now review whole releases instead of single pull requests, and lose the right to skip review after outages they cause.

Company story · critique & review · roles · Human approval

03

LeadDev

AI coding rules matter more than your model

Michael Hill · 18 Nov

A benchmark of over 30 AI models found that curated security rule files cut risk in AI-written code by up to 70%, while vague instructions did little.

How-to · artifacts · agents · Agent context files

04

The Beautiful Mess

TBM 452: Contexting vs. Context

John Cutler · 18 Nov

John Cutler argues that experts struggle to write down what they know, so teams racing to give AI agents more context may skip building shared understanding.

Company story · collaboration · artifacts · Agent context files

05

UX Magazine

From Prototype to Pull Request

Jacquelyn Halpern · 18 Nov

Salesforce designer Jacquelyn Halpern says designers at Trailhead now open pull requests that merge into production, and that companies must agree how those changes are reviewed and owned.

Company story · design systems · handoff · After the prototype

Swipe for 02 to 05
19 Nov
The note

As AI agents write most of the code at some companies, human review of every pull request is no longer keeping up. Five pieces show teams moving the human check to approval rules, written instructions, whole releases and pull requests opened by designers, and one warns what written instructions leave out.

Quentin Rousseau, chief technology officer of Rootly, which makes incident management software (19 Nov · 01), studied 3,661 pull requests merged into its main code repository between August and early November 2026. Monthly merged pull requests there rose from 356 to 2,053 in a year, mostly written by AI agents, while only two more engineers approved them. He found that 90.5% were approved without a single human review comment, and 43.3% were approved on an earlier commit than the one merged. He now proposes that Rootly's AI reviewer approve only low-risk changes outside sensitive code, under a rule a person owns. He has not switched it on, and plans first to record what the AI reviewer would approve while people still decide.

Harjot Gill, founder and chief executive of CodeRabbit, a company that sells AI code review (19 Nov · 02), described a more radical setup to Gergely Orosz of The Pragmatic Engineer. Gill says he shipped over 5,000 pull requests in one month with AI agents. CodeRabbit's engineers now review releases that bundle several pull requests, instead of each pull request on its own. Engineers on a "review exempt list" push to production without review, and each outage traced to them is a strike that can end the exemption. CodeRabbit sells AI review, so its account of where human review is heading is also its business case.

Michael Hill, managing editor of LeadDev, a publication for engineering leaders (19 Nov · 03), reported a benchmark by Secure Code Warrior, a company that sells secure-coding training. It ran more than 30 AI models through over 15,000 app builds. Curated security rule files given to the AI models cut measured risk by up to 70%, while a vague request for secure code did little. Hill adds LeadDev's own survey: 81% of 591 engineering leaders said AI tools already have write access to their code, yet 32% have no formal policy for AI agents.

John Cutler, a product and organisation coach who writes the newsletter The Beautiful Mess (19 Nov · 04), questions the push to write everything down for AI agents. He observes that experienced people struggle to explain what they know, and people asking for context struggle to say what they need. People usually build shared understanding by working alongside one another. His own team dropped an elaborate effort to give AI agents context once it saw that the team was avoiding building shared context itself.

Jacquelyn Halpern, who works on product design and AI at Trailhead, Salesforce's learning platform (19 Nov · 05), wrote in UX Magazine about designers whose prototypes now become pull requests. Her teams build with the same components as the product, and some of those pull requests are reviewed and merged into production. She argues that a company expecting this must give designers repository access, a way to preview changes and test data. It must also agree what belongs in a pull request, who reviews it and who owns the change once it ships. Halpern warns that designers can generate more work than a team can reasonably review.

What they add up to

At Rootly and CodeRabbit, the human check is moving from reading each change to setting the rules about which changes a person must read. LeadDev's report suggests that those rules, and the instructions given to AI agents, now carry much of the quality and security load. An inference from Cutler's piece is that the people who write those rules should test them against colleagues' judgement, because much of that judgement is never written down. Halpern's account adds a new source of changes to that check: designers' pull requests, which need the same agreed rules on who reviews them and who owns them.

The case against

Two of the five accounts come from companies that sell AI review or secure-coding training, Rootly has not yet run its policy, and Halpern gives no figures for the designers' pull requests she describes. The evidence that rules and release-level review are safe still comes mostly from the companies selling them.

Receipts
Rootly
When should AI be allowed to approve a pull request?Company story · 19 Nov 2026
Quentin Rousseau · 17 November 2026

Rootly's chief technology officer found that 90% of recent pull requests were approved without a human comment, and proposes letting an AI reviewer approve only low-risk changes.

The Pragmatic Engineer
Software factories and code reviews with Harjot GillCompany story · 19 Nov 2026
Gergely Orosz with Harjot Gill · 18 November 2026

CodeRabbit's chief executive Harjot Gill says his engineers now review whole releases instead of single pull requests, and lose the right to skip review after outages they cause.

LeadDev
AI coding rules matter more than your modelHow-to · 19 Nov 2026
Michael Hill · 18 November 2026

A benchmark of over 30 AI models found that curated security rule files cut risk in AI-written code by up to 70%, while vague instructions did little.

The Beautiful Mess
TBM 452: Contexting vs. ContextCompany story · 19 Nov 2026
John Cutler · 18 November 2026

John Cutler argues that experts struggle to write down what they know, so teams racing to give AI agents more context may skip building shared understanding.

UX Magazine
From Prototype to Pull RequestCompany story · 19 Nov 2026
Jacquelyn Halpern · 18 November 2026

Salesforce designer Jacquelyn Halpern says designers at Trailhead now open pull requests that merge into production, and that companies must agree how those changes are reviewed and owned.