A third of logged GitHub Agentic Workflows runs ended in failure, study of 4.1 million runs finds

GitHub Agentic Workflows, which became generally available on October 28, let teams write automations as Markdown files that a compiler turns into standard GitHub Actions workflows driven by coding agents. A research team published GH-AWTrace, an archive covering 25,014 verified workflows across 5,051 repositories and 4.1 million run records. Of the 562,877 nonempty execution logs they retained, 31.7% were linked to runs with a failure conclusion. The authors caution that failures are not attributable to the agent alone, since runner crashes, provider outages, or genuinely blocked tasks can also cause them. Because public logs are deleted after 90 days, the dataset reflects what survived retention rather than a clean random sample.