GitHub's Former CEO Built 'GitHub for AI Agents': Trails Replace Pull Requests
GitHub's former boss built a code hosting platform that assumes the people writing code are no longer only people. It may be the year's most direct act of self-negation.
Three key facts
Former GitHub CEO Thomas Dohmke officially launched Entire, positioned as "the first Git forge built specifically for agent-human collaboration." Per the company blog, Entire rests on three primitives: the Trail (a review interface replacing the pull request), distributed Git hosting, and Continuous Memory.
The problem Trails targets is stated bluntly: the PR has become "an audit log that only the agent reads." The blog's logic: developers and review agents loop back and forth, long descriptions pile into the PR timeline, and the PR stops giving a human the information needed to decide "this can merge." Trails instead surfaces key information at the top of the page as automated evals, rescored on every push, so a person can judge at a glance whether enough loops between coding agent and review agents have happened to reach a stable state.
The other two pieces are infrastructure rebuilds. Distributed Git hosting claims push speeds up to 89x faster than GitHub and 6x faster than Cursor Origin, with multi-region replication for data residency; Continuous Memory syncs sessions and prompts into the repository so coding agents can search them. For CI, it partners initially with Buildkite and Depot.
Wang Dou's Take
Start with that self-negation — "the PR has become an audit log that only the agent reads." It's a sharp and honest line: over the past two years we stuffed AI into the PR flow, and the result wasn't people getting faster — it was people getting drowned. Agents writing descriptions, agents reviewing, agents replying to agents; the timeline keeps growing while the judgment a human actually needs — "can this merge?" — gets harder to reach. Entire's insight is right: the problem isn't that the tools aren't smart enough, it's that they were designed for a world where humans write and humans review, and we jammed machine-writes/machine-reviews into it.
The second thing worth pondering is "89x." Numbers like that are common in infrastructure and essentially un-reproducible in real workflows — your bottleneck is usually CI queueing, dependency downloads, approval waits, not push bandwidth. So the number's real function isn't telling you "you'll be 89x faster," it's telling you "people in this market genuinely pay for speed."
The most interesting piece is Continuous Memory. Syncing sessions and prompts into the repo so agents can search them sounds natural, but it effectively turns "AI context" into a versionable, auditable, inheritable asset. The end of that road is: agent memory becomes part of the codebase, just like application code. If that holds, then when you inherit a project, what you'll read isn't just code and commit messages — it's also the thinking traces left by your predecessor's agents.
Of course, all of this rests on an unproven premise: that agent-written code becomes the long-term mainstream. If AI coding reverts to "human-led, machine-assisted" in two years, Entire will have paved a road for a world that never arrived.
Source: Entire Blog · The Verge