Entire launches agent-first Git forge with eval-scored Trails and multi-region hosting

Entire, a startup led by co-founder and CEO Thomas Dohmke, has launched what it calls the first Git forge built specifically for agent-human collaboration. According to the company's announcement, the platform is organised around three primitives: a review model called the Trail that replaces the pull request, distributed Git hosting with multi-region replication, and Continuous Memory, which stores agent sessions and prompts inside the repository.

Trails replace the pull request

Entire argues that the classic pull request no longer serves developers in agent-heavy workflows. With coding agents and review agents exchanging long messages in a linear timeline, the PR has become a transcript of machine-to-machine conversation that a human cannot easily use to decide whether a change is safe to merge.

The Trail is Entire's alternative: a dashboard that puts eval scores at the top of the page, recalculated on every push. A Trail can start from a bug report, a feature request, or directly from an agent before or after implementation begins, and a single Trail can group related changes across multiple repositories.

When code is pushed, review agents run independently in isolated cloud sandboxes, each submitting a score and a rationale — an arrangement Entire likens to a panel of judges scoring a skating performance. Across its own repositories the company uses evals for Security, Slop and Human Review Required; if a score is not green, the developer can instruct the coding agent to keep iterating. The interface shows the current run plus the last five runs for each eval, on the reasoning that one-shot agent loops often miss critical issues and that scores should stabilise before merging.

Entire says projects ship with pre-defined agents for code quality, security and slop that can be enabled per repository, and teams can configure their own agents through the UI or a JSON file in the repo. Under the hood the agents run Claude Code and Codex with the latest models, on sandboxes provided by Daytona and E2B. Agents also file findings with severity levels, which can be addressed or dismissed by agents or humans, and per-repo Gates can require that all findings are closed — along with human approval and passing CI checks — before a merge is allowed.

Distributed hosting and claimed speed

Rather than building on an existing forge, Entire wrote its own source control server: a distributed object store paired with a transactional ref database that remains compatible with Git. Repositories can be hosted natively on Entire or mirrored from GitHub, replicated across regions, or pinned to a single region for data residency. The service is live in the US, EU, Australia and India.

Entire claims this makes it the fastest Git forge in production today. Citing its own open-source benchmarks, the company reports pushes up to 89x faster than GitHub and 6x faster than Cursor Origin, with 418 pushes per second for Entire versus 65 for Cursor, 15 for Pierre, 10 for GitLab and 4.7 for GitHub. These are self-reported figures from benchmarks the company publishes itself.

Continuous Memory and CI partnerships

The third primitive, Continuous Memory, synchronises agent sessions and prompts into the Git repository and keeps that data constantly available and searchable to coding agents — an attempt to make agent context durable and portable rather than trapped in a single editor session.

For continuous integration, Entire is partnering rather than building, positioning itself as an open ecosystem. Buildkite and Depot are the launch integrations. CI is configured at the project level, since multiple repositories in a project often share infrastructure; once Buildkite is connected, every push builds on Buildkite's or the customer's own infrastructure, and a failing build blocks merging. Agents can drive the entire loop themselves via Buildkite's API, CLI or MCP, and GitHub Actions users can migrate workflows with bk pipeline convert.

Depot, whose infrastructure is compatible with GitHub Actions, runs existing workflows without GitHub in the middle. Its depot ci migrate command discovers GitHub Actions workflows, promotes them to Depot CI and flags anything that needs review, with companion commands for importing secrets and variables. Depot bills by the second, so faster jobs cost less.

Why it matters

This is a full-stack bet that the pull request, long the centre of code review on hosted forges, is the wrong abstraction once agents write and review most of the code. Entire's wager is that the human's job shifts from reading diffs to judging confidence: watching eval scores stabilise across agent iterations until a merge looks safe. Distributed hosting and Continuous Memory target the other agent-era frictions of throughput and context. The caveats are real — the performance numbers are the company's own, review quality depends on the same class of models doing the coding, and Entire is asking teams to leave GitHub's ecosystem at a moment when incumbents are also adding agent features — but as a coherent alternative to the PR model, it is one of the clearest articulations yet of what an agent-native forge might look like.