Vibeleaderboard
← All Intel
Intel / article

Catching Developers in the Flow: Low-Latency Agentic Program Repair at Google Scale

Source
arxiv.org
Author
Celal Ziftci, Spencer Greene, Ray Liu, Livio Dalloro, Lorenzo Dini
Date

Context

When a change fails tests before being submitted, developers lose time and . A paper from Google engineers, accepted at ASE 2026, describes FlowAgent, an agent deployed at Google that proposes fixes inside its internal review and editing tools. It generates and validates candidate fixes in a loop and uses filters before and after execution to skip low-quality suggestions under tight limits.

In a manual of 195 real failures, 67.18 percent of its suggested fixes were correct. After company-wide deployment it suggested fixes on 295,508 changes; developers previewed 65,069 and applied 28,554. Applied suggestions measure use, not independent correctness. The authors say interviews found developers received it well while challenges remain.

Key takeaways · AI-distilled
  • Most automated repair work targets post-submit, offline fixing; FlowAgent targets pre-submit CI failures, where suggestions must arrive fast enough to reach developers before they switch context.
  • In a manual evaluation of 195 real-world test failures, FlowAgent suggested a correct fix 67.18% of the time.
  • After Google-wide deployment it suggested fixes on 295,508 changes; developers previewed 65,069 and applied 28,554, roughly 44% of previewed suggestions.
  • Developer interviews found the suggestions useful and autonomous repair in the workflow well received, while the authors say challenges remain. The paper is accepted at ASE 2026.
Terms in this piece · Glossary
  • Context and context window — Context is the information available for a generation; the context window is the capacity available to hold it.
  • Latency — The time between asking for a result and having it, including the model's generation time, tool calls, retries and any queueing.
  • eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
  • Program — A set of instructions stored so that a computer can run it.
Recommended reads
Comments

Checking sign-in…

Loading comments…