Fast Code, Slow Feedback: Fixing CI for the Age of AI Agents
DevOps.com, Tuesday, November 3rd, 2026
A Harness engineer explains five ways to speed up CI as AI coding agents multiply pull requests, pipeline runs and compute costs.
Hemanth Kumar Mantri, a principal engineer at Harness, writes that agents now produce code faster than CI can validate it, citing GitHub data that over 40% of pull requests began with an agent draft by September 2026.
He recommends shallow or blobless checkouts, dependency caches keyed on lockfiles and scoped against poisoning, Dockerfiles ordered for layer reuse, test selection based on the build graph, and quarantining flaky tests instead of blanket retries.
Each fix comes with a caveat, such as needing full-suite fallbacks for test selection.
He advises measuring speed and reliability together, including cache hit rate, cost per merged change and escaped defects.