Greptile co-founder Daksh Gupta traces the shift from code completion to agents that create whole pull requestsAn AI coding agent is a tool-using AI system that can inspect, modify, and validate software within a repository.. He says his company reviews more than a million pull requests each month and used author fields, pull request descriptions and branch names to estimate which were largely or entirely AI-generatedAI-generated code is software source or configuration produced substantially by an AI model from prompts, context, examples, or tool feedback.. By those signals, about a quarter of the reviewed pull requests were AI-generated, up from less than one percent early in the previous year.
The comparison found no large difference between AI-generated and human pull requests in revert rates, Greptile's reported high-severity review findings or rounds of review before merge. The talk also shows that failure types can differ: Daksh Gupta reports variation across agents in findings such as SQL injection. These are observations from Greptile's customer data, and the transcript does not establish that the groups had equally difficult tasks or provide statistical uncertainty for the comparisonsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
Daksh Gupta argues that rapidly rising pull request volume calls for stronger validation before mergeCode review examines proposed source changes to assess their correctness, clarity and suitability before or around integration.. His proposed checks ask whether a change breaks the user contract, increases the chance of a future break, and fulfills its stated intentAgent completion verification checks evidence that an agent achieved the requested outcome rather than merely producing intermediate work.. He describes sandbox execution, mocked inputs and browser agents as ways to test those questions.
Watch on YouTube
