SWE-Touch Benchmarks Coding Agents When Users Edit Mid-Run

A benchmark from the Chinese Academy of Sciences tests something every existing coding-agent evaluation assumes away: what happens when the user modifies source files while the agent is still running. Standard harnesses feed a task in and collect a patch out, undisturbed. Real usage is not like that — developers edit alongside the agent constantly, and whether it notices, clobbers the change or reconciles it is untested. This is a plausible part of the gap between strong benchmark numbers and the sense that agents are less reliable in daily use than their scores imply. Worth watching as vendors start reporting against it.