Agent-Native Mobile Testing

Agentic CLIs are converging on a single substrate for mobile testing — MCP servers expose vendor capabilities, AGENTS.md briefs cross-runtime behavior, hooks enforce deterministic invariants, OpenTelemetry observes the result. The vendors building real-device clouds for AI-driven testing are the first ones who have to solve this in production.

This feature works through what plugin authoring looks like in that world, using Kobiton’s published kobiton/automate plugin as the running case study. Three chapters, expanding as the series ships:

• Chapter 1 — AGENTS.md as a Cross-Tool Plugin Brief (May 11): a 5-device parity sweep, the gaps it exposed, and the operational facts a good AGENTS.md would close.
• Chapter 2 — Making Agents Reliable on Real-Device Clouds (Jun 9): the partial-failure seams between MCP tool calls, and the agents + hooks + documented limitations that close the Chapter 1 gaps. Published on Kobiton’s engineering blog.
• Chapter 3: Text-first AI triage on session logs. (in progress)

Each chapter cites and links back to where it was first published; the chapter form here adds connective tissue that didn’t fit the original post format.