Key Takeaways

  • Jay Parikh, EVP of CoreAI at Microsoft, skips pull requests and issue trackers entirely when building with autonomous coding agents in his own repositories.
  • GitHub is designing its core building blocks to be configurable, accommodating both solo agent-driven creation and strict enterprise review cycles.
  • The platform supports 225 million developers, spanning non-technical creators sharing prompt instructions and teams managing legacy Java and .NET systems.
  • Modern software repositories increasingly store plain markdown files and prompt instructions rather than traditional compiled syntax.

Jay Parikh Skips His Own Company's Review Gates

Pull requests have been the standard unit of software collaboration for fifteen years. A developer writes code on a branch, opens a ticket, waits for a colleague to approve the diff, and merges it into the main branch. When autonomous agents write the code in seconds, that gate turns into friction.

Parikh admits that he discards the ceremony when building his own projects. “There's repos I have that I tinker with and play with,” (play from 7:38) Parikh said. “I don't do PRs right and to me I just like I'm just constantly just firing in ideas and I'm just like building building building building and I don't want to deal with issues or PRs anything.” (play from 7:41)

When a human works directly with an agent, code generation is rapid and iterative. Opening an issue, branching, submitting a pull request, and reviewing your own automated output creates dead time. For solo builders and rapid prototyping teams, the pull request is no longer an asset; it is an obstacle.

The Platform Dilemma: 225 Million Users and Two Worlds

GitHub cannot simply delete the pull request. The platform operates under a split reality. On one side are new builders who use natural language to direct coding agents. On the other side are risk-averse institutions managing mission-critical backends.

Parikh highlighted this divide: “Now you don't have to be a skilled software engineer to use GitHub, right? So now with coding agents can do a lot of things for you and then you can share an idea.” (play from 5:39) These creators often store markdown files containing system prompts and task instructions instead of standard executable source files.

Yet the old world is not disappearing. “I think that there's still also a lot of code that exists and you think about all of the code that's been written in Java and .NET that's running business-critical massive businesses today,” (play from 8:43) Parikh noted. “Those are all in the process of like you have to make security fixes, you're modernizing them, you're still optimizing them.” (play from 8:58)

Because of that balance, GitHub is treating core tools as modular pieces rather than mandatory checkpoints. “Remember across 225 million developers, big enterprises, small, students, etc., we do have to provide a platform with many concepts,” (play from 7:22) Parikh explained. “Then you should be able to assemble these concepts together and however you want to do work.” (play from 7:33)

Fast Loops Versus Compliance Gates

If you run an early-stage team, enforcing pull requests for every single agent generation slows you down without adding safety. Human code review worked because humans wrote small batches of code slowly. An agent producing hundreds of lines across multiple files creates review fatigue. Human reviewers skim the diff, click approve, and provide an illusion of safety without real scrutiny.

Instead of treating every branch like a multi-million-line Java application, teams should separate automated agent changes into two tracks. Fast-moving exploratory code and prompt adjustments can bypass branch protections and pull requests entirely. Production infrastructure, data schemas, and legacy integrations retain mandatory automated tests and approval gates. Match the workflow to the velocity of the tool.

What to Do With This

Audit your repository branch protections on Monday morning. Split your internal projects into rapid agent sandboxes and protected release branches. Turn off mandatory pull request approvals for solo prototypes and prompt directories, and replace human diff reviews with automated integration tests that run directly on merged commits.