Approve the safe pull requests. Keep people on the rest.
Automerge scores each pull request and approves the safe ones. You start it in Dry run, where it only posts its decision as a comment until you switch it on.
Start for free or Book a demo
Your agents write faster than anyone can review.
Coding agents made a pull request cheap to write. Reviewing one did not get cheaper, so the queue grows, including the routine changes nobody needed to read line by line. Automerge takes that routine share off your reviewers and routes everything else to a person.
Pins the commit
Every evaluation runs on one fixed commit. A new push means a new evaluation, and approvals on older commits are dismissed.
Holds what you protect
A pull request that touches a protected path goes straight to a person, without being scored.
Scores the change
One AI evaluator per criterion reads the change and scores it from 0 to 100, with a rationale you can read.
Approves or holds
It approves only when the combined score clears your threshold and no criterion falls below its floor. Everything else waits for a reviewer.
Three questions every reviewer asks.
Every repository starts with three criteria. Change their weights and prompts, add your own, or set a floor that rejects a pull request on its own, whatever the combined score.
How far could a failure reach?
How far a failure could reach, and how easily the change can be rolled back.
Is there proof it works?
Tests, passing checks and reviewer signal that the change is correct.
Is this routine?
How routine the change is, from a mechanical edit to new load-bearing code.
Autonomy you grant one step at a time.
You start each repository in Dry run, where Automerge comments what it would do and changes nothing. Move a repository up once its decisions match your reviewers' decisions, or give more autonomy only to the paths that have earned it.
Nothing runs
Takes no action on this repository.
Watch it decide
Comments its decision without approving or merging.
Approve, a person merges
Approves qualifying pull requests. Never merges.
Approve and merge
Approves and merges qualifying pull requests.
Check its decisions before you trust them.
Every decision, score and rationale lands in Dash0 next to what your reviewers actually did, so you know when a repository is ready for more autonomy.
Graded against your reviewers
Each decision is graded as a correct approval, an incorrect approval, a correct hold or an unnecessary hold, so you can see where the bot and your team disagree.
Replay before you change
Test a new threshold, prompt or criterion against a fixed set of past pull requests and compare it with what is running. Tests never change what is running.
Was this decision right?
Reviewers mark a decision right or wrong, with a reason, from the pull request in Dash0 or from the bot's comment on GitHub.
Built to fail closed.
You can only rely on an approval agent if you can predict when it will hold back. If an evaluator fails, a path is protected or a score falls short, Automerge leaves the pull request to a person.
Protect what matters most
Mark migrations, infrastructure or billing as Protected and they always go to a person. A pull request runs at the least autonomous level of any file it touches, and renamed files are traced to where they came from.
When in doubt, it holds
If an evaluator fails, nothing is approved. In Auto-approve and Auto-merge, a commit whose CI is failing is never approved.
A pull request can't rewrite its own rules
Rules live in Dash0 and steering files are read from the base branch. Titles and descriptions are treated as signals, not proof.
Advisory, never in the way
Its GitHub check is advisory and blocks nothing unless you make it required. It never re-approves a pull request a reviewer dismissed, and only admins can change a repository's configuration.
Reads what a reviewer would read.
Each criterion chooses its own read-only tools. Evaluators read tracked files at the evaluated commit and never run your code.
The code around the change
Read files, grep, list files, git log, blame and the diff against the merge base.
The change and its signals
Metadata, changed files, CI check runs, reviews and comments, and commit signatures.
Who wrote it, and how it went
The author's familiarity with the files, their delivery record, AI review agent activity, and the repository's evaluation history.
How the change was built
Which harness, tools and skills produced the branch, read from AI SDLC Insights.
Start in Dry run.
Connect GitHub, pick your repositories, and see what Automerge would have approved before it approves anything.