Pull requests remain important for team review, but they need not be mandatory for every project, Jay Parikh said in a November 7, 2026, Latent Space interview. Parikh, Microsoft’s CoreAI executive vice president and GitHub’s leader, said he sometimes turns an idea directly into code in personal experimental projects without opening a pull request or managing issues. Larger organizations, by contrast, may still need reviews, permissions and change records. He said GitHub could consider making workflows configurable rather than requiring the same process for everyone.
The question has become more pressing as coding agents take on more work. According to GitHub figures cited in the report, agents drafted the first version of more than 40% of pull requests by September 2026, up from 17% in March. The number of agent-initiated pull requests grew roughly sixfold over that period.
Parikh also described GitHub’s approach to model selection. Copilot Harness connects models and tools and coordinates development work, while Copilot Fusion aims to choose models according to the task. More model calls can mean more time and higher AI-credit use, so GitHub wants to balance quality, speed and cost while leaving developers a choice of models and tools. In a comparison GitHub presented from the Artificial Analysis Coding Agent Index, Fusion Max and Opus 5.5 Max differed by less than two points; GitHub said Fusion Max cost about half as much per task and took about 40% less time.
For security, GitHub presented Copilot code review Max as a way to identify harder-to-spot code problems and demonstrated a push-time block on an accidentally exposed secret. Parikh said GitHub is also working with Microsoft’s security team and open-source maintainers on package-manager risks and compromised repositories.
Replacing open-source dependencies with AI-written code would not eliminate maintenance, Parikh cautioned. Responding to a hypothetical project that reduced 2,000 dependencies to 20, he noted that the team would then have to maintain the functions previously supplied by the other 1,980 components.
Parikh said organizations need to assess coding agents against outcomes, not just token spending. He mentioned revenue, cost savings, user engagement, retention and customer satisfaction as possible measures. On GitHub’s own infrastructure, he said the platform is adjusting to increased automated activity; a GitHub Universe 2026 article by Parikh projected completion of its move to a multi-region Azure deployment in early 2027.