Your Biggest User Isn't Human: Designing DevTools for AI Agents

Mokos describes Expo’s shift from building its own coding harness to supporting the agents developers already use, such as Codex, Claude Code, and OpenCode. He says about 50% of commands across Expo’s and EAS’s CLIs are already run by coding agents.

Agents need to write code, build apps, and test them, but iOS development adds requirements such as macOS and Xcode. Expo’s cloud simulator service, then in beta, lets agents test apps remotely; Expo also provides tools for building apps and delivering them to TestFlight and the App Store.

Expo can see how agents use its CLIs, documentation, skills, MCP servers, and simulator, but it cannot see where they struggle inside third-party coding harnesses. Mokos says internal agent telemetry helped, though its sample was small. Expo then added feedback instructions to its skills and Markdown documentation, receiving dozens of useful reports per day, including conflicting guidance and an animation snippet that crashed apps.

Expo routes submissions through a classifier to filter spam, then uses embeddings and a frontier model to match feedback with existing Linear issues or create new ones. Background agents fetch issues, investigate them using simulators, and open pull requests for human review. Mokos says agents can also propose eval tasks by describing what they tried, what they expected, and what happened.

Mokos recommends giving agents a way to report problems, gathering conversation history where possible, and turning observed struggles into evals. The aim is to use agent feedback to improve tools and keep the feedback, fixes, and evaluation cycle moving.