Entire Wants to Replace the Pull Request. I Read Its 89x Benchmark.

Pull requests were built for humans taking turns, and GitHub's former CEO wants to replace them. I read Entire's launch post, docs and the benchmark behind its "89x faster than GitHub" headline, so here's what that number measures.

The bottleneck

I run a lot of coding agents in parallel, and they write code fast. My slow step now is the pull request. Every agent opens one, and every one waits for a human to read it. PRs were designed for people taking turns. My agents work all at once.

The data points the same way:

• A Harvard working paper by Chen and Stratton looked at 718 companies. After they adopted coding agents, the time from pull request to merge went up 49%, about 3.45 extra days on a 7-day baseline. The paper is a working paper and has not been peer reviewed. Its data comes from Jellyfish, an engineering-analytics firm.
• LinearB, a metrics vendor, found agent pull requests wait about five times longer before anyone picks them up: 17.6 hours versus 3.4 hours for unassisted PRs, at the 75th percentile.
• Anthropic went from about 2,500 merges a day in May to over 30,000 a day in September. It says the bottleneck now is "all of the developer infrastructure built for human scale: source control, builds, and CI." FIRST-PARTY · ANTHROPIC Anthropic also notes its own workload is unusual.

So when the former CEO of GitHub sets out to replace the pull request, I pay attention.

Who built this

Thomas Dohmke was GitHub's CEO from late 2021 until he left in 2025. Copilot grew up on his watch.

In February 2026, his new company, Entire, raised a $60 million seed round led by Felicis, at a $300 million valuation. VENDOR · ENTIRE Felicis, the lead investor, calls it the largest seed round in developer tools history.

On October 26, Entire launched its own Git forge, two days before GitHub Universe (October 28-29 in San Francisco). VENDOR · ENTIRE

His pitch, in his words: the pull request "has become an audit log that only the agent reads." VENDOR · ENTIRE

Entire's answer has three parts, and they map to the next three sections:

1. Trails
2. Continuous Memory
3. Distributed Git hosting

What Trails actually does

Dohmke describes a pull request as a linear timeline. Comments stack up, CI reports a status at the bottom, and the merge button waits under all of it.

A Trail flips that layout and puts eval scores at the top of the page. VENDOR · ENTIRE Here is how it works:

• It can start before any code exists. You write the intent first, as a bug report or a feature request, or your agent opens a Trail when it's ready to push.
• One Trail can span several repos.
• Review agents run on every push. The agents you configure each run in their own cloud sandbox and return a score and a rationale. VENDOR · ENTIRE On its own repos, Entire runs Security, Slop and Human Review Required. The presets it ships are code review, security review and Slop.
• Humans read the scores. Dohmke's analogy: "Think about it like a jury judging an ice skating performance."
• Each score shows history. You see the current run plus the last five. If slop is orange, you tell your agent to keep pushing fixes until it turns green.
• Findings point at code. Each finding carries a file and line location, and agents or humans address or dismiss it.
• Gates decide when you can merge. Human approval, passing CI, findings cleared, branch up to date. You set the rules per repo. VENDOR · ENTIRE

For me, that's the shift. I check whether the loop has settled, then read the lines the scores point to.

Continuous Memory

Most agent sessions start from zero. Entire's framing: a new context window "has to re-learn everything which is not only slow but also expensive." VENDOR · ENTIRE

Its fix starts with the open-source Entire CLI, MIT licensed. It saves each agent session next to the commit it produced: prompts, tool calls, files touched, all stored in Git. VENDOR · ENTIRE It redacts the secrets it detects before writing a checkpoint. The CLI's own README calls that redaction best effort, and if your repo is public, those transcripts are public too.

Then come two new open-source tools:

• entire-graph maps your codebase locally with tree-sitter and makes no network calls. It covers definitions, callers, types, routes and change impact. Entire says it cut token use by up to 71% in its own testing. VENDOR · ENTIRE No methodology for that figure is published.
• entire-brain indexes past sessions and ties them to that map. Entire describes the result as giving the next agent the "why" behind the code, plus what tests broke the last time someone changed it.

Their line: "The repository is the only memory your agent should need."

The speed claim

Entire says it pushes up to 89 times faster than GitHub, and six times faster than Cursor's Origin. VENDOR · ENTIRE This is Entire's claim. I ran no benchmark of my own.

Here's the math. Entire's chart shows:

Forge	Pushes per second (Entire's chart)
Entire	418
Cursor Origin	65
GitHub	4.7

418 divided by 4.7 is 88.9. That's the 89.

The chart is measured with ForgeMark, Entire's own open-source benchmark (MIT). Its headline test is many agents pushing to one repo at once, each on its own branch. VENDOR · ENTIRE

Now the footnotes:

1. GitHub may be throttled. ForgeMark's own README warns that github.com "applies secondary rate limits and abuse detection to high-volume content writes," and that high-volume automated load testing of github.com isn't sanctioned. So GitHub's own throttling may be capping its number.
2. The settings are unpublished. The chart's concurrency, commit size and region aren't published.
3. The number is native hosting. In mirror mode, every push writes through to GitHub, and the README says mirror mode "can never beat GitHub itself." So the 418 has to come from Entire's native hosting. A mirrored repo still pushes through GitHub.

The design behind the speed is the interesting part. Full copies of your repo can live in each region you pick, so cloud agents clone from nearby. That's live today in the US, EU, India and Australia. VENDOR · ENTIRE

Everyone sees the same bottleneck

• GitHub: in March, GitHub said Copilot was handling more than one in five code reviews on its platform. FIRST-PARTY · GITHUB
• Anthropic: Claude Code Review became generally available on October 15 for Team and Enterprise organizations. It runs on your pull requests, and Anthropic says most reviews now finish in six to ten minutes, down from about twenty. FIRST-PARTY · ANTHROPIC
• Cursor: Cursor opened its own forge, Origin, to paid users in August, after its deal for Graphite, the stacked pull request tool.

All three keep the pull request and add AI review. Entire wants to replace the pull request page itself.

Hacker News is split. One commenter, stryan, wrote: "I struggle to imagine a scenario where I willingly choose the 'Trails' view." Another, philipwhiuk, called it "a high-AI-trust approach" and added, "Personally I'm still trying to get my company to add AI review to MRs."

Fair point. Trails fits teams that already let agents write most of the code. That's my setup. Yours may get there later.

Price and switching

• Free right now: hosting, Trails and the CLI. You pay only when review agents run, at the model provider's published token rates. VENDOR · ENTIRE
• Trial credit: every billing account gets a one-time $100 credit, good for 90 days, with no card required. VENDOR · ENTIRE
• Keep GitHub: you can mirror your repo, keep GitHub as the source of truth, and run Trails on top of your pull requests. On mirrors, GitHub branch rules still apply.
• Full migration leaves things behind: issues, pull requests, release records, permissions and CI stay on GitHub. VENDOR · ENTIRE
• CI: Entire partners with Buildkite and Depot, and both have commands to convert GitHub Actions: bk pipeline convert and depot ci migrate.

Ideas for getting paid

This part is my opinion: two service ideas to try.

Idea one: write review agents for other teams. A Trail agent is, roughly, instructions, a model and a score. A payments team wants a payments reviewer. A team mid-migration wants a migration-risk score. Package those as a setup service.

Idea two: fix the review queue. Measure how long agent pull requests sit before pickup. Set up gates, preset agents and CI. Then show the before and after. The same loop works with Copilot review or Anthropic's Code Review on GitHub.

Price it per repo, or as a monthly retainer. Before you quote, run a few reviews on your own trial credit and track the token cost per review. That's your floor.

The verdict

The pull request was built for humans taking turns. Agents work all at once.

Entire is betting its whole forge on that. Treat the speed numbers as Entire's claim, and try Trails on one mirrored repo before you move anything.