As AI agents take on more of the building, teams are replacing case-by-case checking with one repeatable path that every piece of agent work passes through. Shopify, Dropbox, Microsoft and the consultancy Modus Create each describe such a path: a shared test, a standard task, a recorded change, a single route for problems or a fixed set of human checks.
Thiago Magalhaes (6 Nov · 01) wrote on the engineering blog of Shopify, the commerce software company. He described how its mobile team rebuilt its iOS and Android apps for AI agents that write most of the code. The slowest step had been proving that a screen worked, which needed an emulator, a build and a tap script. The team made every screen publish its state and its allowed actions as data, so a command-line tool can drive the app without a phone. One test file now runs against both apps, and a pull request cannot merge if it passes on only one. More than 1,700 behaviour tests finish in about four minutes.
Pravi Garg, Sourabh Yerfule and Gennadiy Civil (6 Nov · 02) wrote on the engineering blog of Dropbox, the file storage company. They described how engineers built its new AI experience for web and desktop before a November launch. As the work went on, engineers moved from asking AI for help with single coding steps to handing coding agents defined tasks, including implementation and testing. Each task started with a written plan agreed by the engineer and the agent, and the agent recorded its build and test results for an engineer to review. A test cleanup that had been deferred for two quarters was split into 53 tasks and finished in one week.
Daniel Reiter Horn and Rich Chan (6 Nov · 03), engineers at Dropbox, explained the design choices behind the AI agent that now changes customers' files inside the product. For a broad request, the agent must propose a plan, and a person must accept it before the agent may ask to change files. Folder access appears as an Allow or Deny card, and the permission checks run outside the agent's workspace. Dropbox records which conversation produced each file version. A person can therefore undo one conversation's changes and leave other work alone.
Kevin Kotani (6 Nov · 04), a digital experience lead in Microsoft Digital, the company's IT organisation, described how Microsoft made Copilot, its AI assistant, work better in Japanese. Staff in Japan reported that Copilot also had to get names right, follow meetings that switch between Japanese and English, and use polite business Japanese. Bug reports were duplicated across many teams. Microsoft formed one localisation team in Japan that links field staff, the Copilot product group and language experts, and gives each problem a named owner. More than 300 native speakers tested a preview before broad rollout, and name recognition in Japanese meeting transcripts rose from about 40% to 66%.
Flor Cabello (6 Nov · 05), a product leader at Modus Create, a software consultancy, described a five-month rebuild of a large online shop's mobile app, finished before Black Friday. The team moved close to one designer for every developer, because developers working with AI coding agents moved faster. Written instructions for the agent made it propose an implementation plan, which an engineer approved, edited or rejected before any code was written. Before opening a pull request, the engineer compared a screenshot of each screen from an iOS emulator with the approved design. The engineer who submitted a pull request had to understand and defend the code, and testers checked each feature against its original ticket.
Each team found that checking agent work one piece at a time would not keep up, so it standardised the path the work takes. Shopify has one test for two apps, and Dropbox has one plan, test and review format and one undo per conversation. Microsoft has one route and one owner for each language problem, and Modus Create runs every feature through the same plan approval, design comparison and ticket check. Reading these as one pattern is an inference, not a claim any author makes.
All five accounts come from companies describing their own products, internal programmes or delivery method, and none reports a failure that its checks missed. Modus Create notes that testing and code review grew harder to keep up with as the number of pull requests rose. The controls may work less neatly in practice than they read on a company blog.
Shopify rebuilt its mobile apps so AI agents can test every screen without a phone, and one shared test file keeps the iOS and Android apps in step.
Dropbox engineers handed coding agents defined tasks, each starting from a written plan and ending in an engineer's review, to finish a product launch on time.
Two Dropbox engineers turned side projects into the AI agent that changes customers' files, with plan approval, folder permission cards and an undo tied to each conversation.
Microsoft gave Copilot's Japanese-language problems one reporting route with named owners and had native speakers test early; daily use in Japan then peaked at 93%.
Modus Create rebuilt a shop's mobile app in five months with coding agents, moving to nearly one designer per developer and requiring an engineer's approval of each plan before code.




