As AI agents write most of the code, three engineering teams report that the work has moved to what surrounds it. Each team wrote its engineers' judgement down where the agents and the build checks read it every time.
Paul Constantinides, Anubhav Dhoot and Jackie Chambers de Freitas (11 Nov · 01) write from Salesforce, which makes customer relationship management software. They describe how about 500 engineers in its Hyperforce Runtime Platform organisation took up AI. Useful AI work was staying inside single teams, efforts overlapped, and tight delivery dates left little room for a new programme. A steering committee started one self-updating Claude Code plugin for shared instructions and skills, and each team that built a new skill had to support it. The authors report that cycle time fell from 22 days to 11 and sprint delivery rose from 66% to 90% within three months.
Remi Vanhee (11 Nov · 02) explains on Shopify's engineering blog how about 30 engineers rebuilt the Shopify app, which merchants use to run their stores, in 12 weeks. AI agents wrote most of the code, so review became the slowest step, because pairs of engineers could not read every line. The team made each pull request a small piece that a person can read, covering iOS and Android together, and an AI agent reviewed it first. Each evening an agent turned that day's human review comments into rules in the team's skills and architecture file. Vanhee reports that the number of changes requested per pull request kept falling.
Eguchi (11 Nov · 03), an engineer on the LITALICO Nursing Net development team, published an account in Japanese of raising test coverage from 23% to 84% over a year and a half. LITALICO is a Japanese company that runs software and services in education, disability welfare and elder care. Hand-written tests and a new continuous-integration setup had reached 42% before the company rolled out the agent mode of GitHub Copilot, an AI coding assistant. Eguchi then wrote a personal testing routine down as prompts and later as shared skills, and the AI agent produced 50 to 100 test files per pull request. Eguchi kept the trade-offs for people, such as avoiding mocks, and now searches the output for worthless tests, such as tests that expect a server error.
In all three accounts, a person's judgement about what good code looks like became the limited resource once AI agents wrote the code. Each team put that judgement into a file that agents and checks read on every task: a shared plugin, rules updated from review comments, or a testing skill. The inference, which is not any one author's claim, is that senior engineers now spend more of their time writing and maintaining those rules.
All three accounts come from the companies themselves, and none reports what the new rules missed or how many defects reached customers. Coverage, cycle time and review comments can all improve while the quality of what ships stays the same.
Salesforce's Hyperforce Runtime Platform group spread AI use to about 500 engineers through one shared plugin, with each new skill owned by the team that built it.
Shopify rebuilt its merchant app as native code in 12 weeks with AI agents writing most of it, by making every pull request small enough for a person to judge.
A LITALICO engineer raised test coverage from 23% to 84% by writing a personal testing routine into AI agent skills, while keeping the trade-off decisions for people.


