The Orchestrator

If agents write most of the code, review each other’s work, and open their own pull requests, is a version control system designed in 2005 for human collaboration still the right tool? There are people who say Git is outdated for this, and the argument is not silly: Git was built around a human pace of commits, a human review cycle, and a repository model that assumes the person pushing the code can be asked, later, why they pushed it.

The answer at the time of writing is that the agentic era still needs source control — this book was written with it — but the volume assumption has already broken. When one human developer pushes a few times a day, GitHub’s architecture barely notices. When hundreds of agents branch, commit, and open pull requests around the clock, the assumptions underneath the hosting layer fail: consensus-heavy coordination becomes the bottleneck, and review queues designed for one pull request per person per week are suddenly expected to absorb hundreds.

That is why the most interesting infrastructure move of 2026 came from Cursor. In August, Cursor launched Origin, its own Git hosting platform — not a wrapper around GitHub, but a Git forge rebuilt for the assumption that agents are the primary committers.¹ Beneath it sits a storage system Cursor calls Continuity: every push lands as an append-only write-ahead log in S3-compatible object storage first, and the Git repository on local disk acts as a warm cache that can be rebuilt from the log.²

GitHub’s answer, so far, is the opposite move. The incumbent did not rebuild its storage layer; it made agents first-class participants in the workflow it already runs. Assign an issue to GitHub’s Copilot coding agent the same way you would assign it to a teammate, and the agent spins up its own isolated environment, works on its own branch, runs the tests, and opens a pull request for a human to review — an asynchronous colleague that is not allowed to merge its own work.³ The newest move says the conversation itself belongs in the record. Entire, founded by Thomas Dohmke — the former CEO of GitHub, who left in 2025 after four years of leading the company through its Copilot era — raised a sixty-million-dollar seed round — the largest in developer-tools history, by The New Stack’s account — on a focused premise: capture and make searchable the entire conversation that goes into the creation of software.⁴ Its first release, an open-source tool called Checkpoints, records the full agent session behind every commit — the transcript, the prompts, the files touched, the tool calls, the token usage — stores it in Git as first-class versioned data, and makes it searchable through what the company calls an AI contribution graph. Where Git has always preserved what the code looked like and who touched it, Entire argues that when agents write the code, the reviewable artifact is the dialogue that produced it.

An orchestrator does not need to host code on any particular platform. What an orchestrator does need is source control that keeps up with fleet volume, an audit trail of who changed what — and “who” now includes which agent, under whose authority — and a place where the answer to “what changed?” is never more than one command away — increasingly, with the conversation that produced the change attached to it.