Why AI coding agents love type errors

AI coding agents operate in a feedback loop and stop when all signals go quiet — meaning bugs that don't surface during the loop get shipped. Type errors are the most valuable signal because they fire at compile time, point to a specific line, and are cheap to produce. Most backend infrastructure mistakes (wrong event payloads, cross-service call mismatches, cron double-firing, schema drift) are 'stringly-typed' and produce no compile-time signal, so agents miss them entirely. Typed infrastructure primitives — like a Topic<T> instead of a queue name string — convert these silent runtime failures into compile errors the agent can catch and fix before finishing. Encore is presented as an example framework that provides such typed primitives, backed by a Rust static analyzer that builds a full application graph. The post also addresses the escape hatch of 'as any' or '@ts-ignore', arguing that suppressing a type error is at least a visible, reviewable decision rather than a silent accident.