Entire Replaces the Pull Request With Agent-Scored Evals

Thomas Dohmke spent years running GitHub, the company that made the pull request the centre of the software world. On Monday he launched a competitor built on the bet that the pull request is now the thing slowing everyone down.

Entire, his startup backed by a $60 million seed round, opened to every developer today. It is a Git forge, meaning it hosts repositories and coordinates changes like GitHub does, but it is designed around a different assumption: that coding agents write most of the code and humans approve it in bulk. Instead of a pull request thread with a diff and a comment stream, Entire ships "Trails" - a review surface where multiple agents score each push against configurable evals, and the developer merges when the scores stabilise.

What Entire actually shipped

The Trail is the centrepiece. When code lands on a Trail, Entire spins up review agents in isolated cloud sandboxes - the company says it runs Claude Code and Codex under the hood on Daytona and E2B infrastructure - and each agent returns a score plus a rationale, the way a judging panel scores a routine. Entire's own repositories run three default evals: Security, Slop, and Human Review Required. A score turning orange is a signal to the coding agent to keep iterating and pushing; the interface shows the last five runs alongside the current one, because, as Dohmke puts it, "one-shot loops are often missing a critical issue."

Findings carry severity levels and suggested fixes. Resolved items hide, open ones stay visible, and per-repo Gates can block a merge until findings are closed, CI passes, and any required human approvals arrive. The entire loop is exposed through an open-source CLI, so an agent can read eval results and continue correcting without a human copy-pasting feedback into a chat window.

The second primitive is distributed hosting. Entire runs a new Git-compatible source control server on a distributed object store with a transactional ref database. The company publishes benchmarks claiming pushes up to 89x faster than GitHub and 6x faster than Cursor Origin, with regional hosting in the US, EU, India, and Australia, plus mirroring for teams that want GitHub to stay the source of truth. The third is Continuous Memory: two open-source, local-first plugins. entire-graph precomputes a tree-sitter map of the repository - definitions, callers, types, routes, change impact - and claims up to 71% token savings. entire-brain indexes past agent sessions into that graph so the "why" behind a change survives the context window.

CI arrives through Buildkite and Depot, with migration commands that convert existing GitHub Actions workflows rather than asking teams to rewrite them.

Why the pull request is the bottleneck

Dohmke's argument is not aesthetic. It is backed by numbers that have been piling up all year. A Harvard study published this month, drawing on 300 million work events across more than 700 firms, found that introducing coding agents raised lines of code by 30%, commits by 20%, and pull requests by 23% - while the resolution rate for actual tracked issues and epics did not move in any statistically significant way. The efficiency gain was absorbed downstream: average review time between a pull request opening and merging ballooned by roughly half, reviewers left more comments, and revisions became more likely.

That is the precise failure mode Trails is aimed at. In the pull request model, the diff is the artefact and the human is the reviewer. In the Trail model, the eval history is the artefact and the human is the approver. Whether that is better or merely faster is the real question, because a scoring agent is not a guarantee - it is a second model agreeing with the first one, and the Slop eval exists precisely because the first model produces slop at volume.

The infrastructure pressure is real too. GitHub has spent 2026 firefighting: its own July availability report logged eight incidents of degraded performance, with 113,930 pull request creation attempts impacted across 50,904 users in a single month, and the company has been accelerating its migration of Actions to Azure to absorb spikes. Reliability, not features, is where the incumbents are exposed.

What this means for builders

If you run agents that write code, four things in this launch are worth stealing regardless of whether you switch forges.

Treat the agent trace as a reviewable artifact. Entire stores sessions, prompts, and tool calls in Git history next to the code. Most teams throw that away and then wonder why the next agent repeats the same mistake.

Score every push, not every merge. Evals that run on each commit and display a trend line give you the signal a diff cannot: is this stabilising or thrashing? A single green run after a dozen orange ones is not confidence.

Put your repository's structure on disk for the agent. A precomputed code graph is cheap compared to re-reading the codebase on every session, and the claimed 71% token reduction is a cost line, not a vanity metric.

Keep the merge gate human-owned. Agent-scored merges are a throughput feature; Gates that require a named human approval for sensitive paths are the control that keeps it from becoming an audit problem later.

The competitive question is not whether Entire beats GitHub at hosting. It is whether "merge without reading every line" becomes a normal engineering practice. Every review tool vendor is now selling exactly that, and the study above suggests the bottleneck has not disappeared - it has been pushed to a new place, where the reviewer's job is to trust a score.

Entire is available today, and the CLI plugins are open source, which means the interesting part is testable this week rather than next quarter.