AI Coding Agents Still Fail Users: Study of 20,574 Sessions Reveals 7 Failure Patterns

Researchers from multiple universities conducted an observational study of 20,574 coding-agent sessions from 1,639 repositories, spanning both IDE and CLI workflows. They operationalized 'misalignment' as breakdowns made visible through developer pushback, then annotated each episode along four axes: form, cause, cost, and resolution. The analysis revealed seven recurring forms of failure, including how agents read projects, interpret intent, follow rules, bound actions, implement code, and report progress.

Key findings show that 90.50% of misalignment episodes impose effort and trust costs rather than irreversible system damage—yet 91.49% still require explicit user correction to resolve. Patterns also differ across IDE and CLI settings, persist across adjacent sessions, and shift over time: while overall failure rates decline, constraint violations and inaccurate self-reporting grow in share. These results highlight fundamental gaps in how AI coding agents understand developer workflows and suggest that current benchmarks fail to capture real-world misalignment experiences.