GitHub is the wrong shape for this new world

Kyle Galbraith, who runs Depot, has a simple but pointed argument: the branch-PR-review-merge loop that GitHub popularized wasn't built for a world where anyone, including a sales rep or an AI agent, can generate a working feature from five sentences of prompt. That workflow made sense when humans wrote code at human speed. It's straining badly now that models routinely produce correct code on the first pass and teams are drowning in parallel branches, more CI runs, and pull requests nobody has time to review properly.

Galbraith's framing is that GitHub, and really any git host, is a collaboration tool dressed up as infrastructure. It's fine for humans debating ideas in comment threads, the same way Google Docs works for editing a blog post together. But once code stops being something a small team writes and starts being something an entire company, plus a fleet of agents, generates continuously, the bottleneck isn't the coding anymore. It's source control, CI, security scanning, and deployment pipelines built around the assumption that a person is sitting there clicking merge.

He's not just griping about slow builds, though that's clearly personal for him as a self-described platform engineer who despises them. His bigger claim is that basic questions are now unanswered at most companies: who owns AI-generated code, who's accountable when it ships, how do you review a volume of changes no human team can realistically read, and how do you weigh token costs against the value of what gets produced. Ask five engineers those questions, he says, and you'll get five different answers, which he takes as proof the old assumptions have quietly stopped holding.

His proposed fix borrows straight from the cloud-native playbook. Just as compute, storage, and networking became reusable primitives instead of individual servers you provisioned by hand, he wants software delivery broken into primitives too: immutable source history, isolated execution for testing changes, reproducible artifacts, caching for deterministic work, identity for tracking who or what wrote the code, and machine-enforceable policy for security and compliance. None of that, he argues, is a git host feature. It's infrastructure that needs to run at machine scale, not meeting-room scale.

Galbraith's bet, unsurprisingly given that Depot sells build and CI infrastructure, is that the next decade's winners won't be the team that ships a nicer pull request UI. They'll be whoever builds the plumbing that lets code get generated, validated, and deployed without a human standing at every gate.