We gave AI agents a mobile dev environment

Shopify gave its AI agent remote Macs to build, test and record mobile app fixes, and the agent's share of merged mobile pull requests rose from 6% to 61%.

Shopify's mobile engineers ran several AI agent tasks at once, but three or four were enough to overload a laptop, and closing the lid stopped the work. River, Shopify's AI agent in Slack, could edit iOS code on Linux but could not build or test the app. Michael Garfinkle, Mauricio de Meirelles and David Cortes describe how the team built Remote Dev, a pool of prepared remote Macs, to fix this.

• Each agent session gets its own remote Mac, and keeping Macs ready in advance cut setup from more than 30 minutes to a few minutes.
• For a bug fix, the agent records the broken behaviour, applies the fix and repeats the same steps, so the reviewer sees the result before and after.
• A live stream of the running app on a web page lets designers and product managers try a change without setting anything up.
• A fast, low-cost AI model tags every agent conversation by outcome, and the team uses the failures to improve the agent's instructions.

An engineering team can ask its AI agents for visible proof that a change works, not only passing builds. The authors also report that adoption needed internal promotion as well as better tools.

Derived by Working Surface from the article; more in the Playbook. Source line: Just because the code compiles doesn't mean the feature behaves as expected.

25 October 2026: AI agents and people outside engineering can change real products safely once a team has built shared templates, written rules and automated checks.