AgentGate: a merge queue that only ships agent PRs it can prove are safe
The problem
Coding agents now write most of the code at real teams — one two-person startup merged 11,040 PRs in a year with 82% of commits agent-signed. Agents report success while breaking things: one developer's release train froze for six days over a double bracket while every run said done, and another team's agent wrote an idempotency race condition with excellent documentation. Existing review tools read the diff; they cannot see which agent, harness, skill version, or prompt produced it, so the same mistake ships again next week.
The idea
Why now
StarSling made CI self-improving, SkillEvaluator grades agent skills with quality gates, t3code gives one control surface across harnesses, and agent-authored code just crossed the majority of commits at real teams. The provenance, grading, and control pieces all landed in the last week.
What it combines
StarSling's self-improving CI + SkillEvaluator's skill grading + t3code's cross-harness control surface + Pulsar's agent-to-agent messaging. CI alone reruns tests, grading alone scores skills, and a control surface alone watches agents — together they are a gate that knows who wrote the code, how good that writer's skills are, and what this repo has already learned not to accept.
MVP
Weekend MVP: a GitHub Action that (1) labels PRs by author agent from commit trailers, (2) runs a skill-grade check on the diff via an existing grader, (3) blocks merge on a configurable risk score with a one-line verdict. Deliberately skip: the self-improving incident loop, cross-harness messaging, and auto-merge — ship the label and block first.
How it makes money
The moat
CodeRabbit, Qodo and Greptile would have to rebuild around agent provenance — which agent, harness, skill version and prompt produced each diff — and around per-repo incident corpora. Their product is a diff reviewer; the gate is a harness-layer system of record, and the per-repo mistake corpus is a data advantage that compounds with every incident a team records.
Distribution
B2B via a GitHub Marketplace app and CI integration; sell to platform-engineering leads at teams where agents already write most code. Land through harness vendors and control surfaces so every agent fleet gets the gate by default.
Why it wins
CodeRabbit, Qodo and Greptile review diffs from humans and agents alike; none of them gate on agent provenance, skill grades, or a repo's own incident history. AgentGate's verdict is 'this agent, with these skills, has made this mistake here before' — a diff reviewer cannot say that.
Risks
Biggest risk: teams distrust an automated gate and route around it. The MVP de-risks it by starting in advisory mode — verdicts posted as comments, merge never blocked — and only enabling blocking after a team sees the gate catch a real bad PR.
Build it with
- StarSling: self-improving CI for agent-volume reposSelf-improving CI: every incident becomes a check that fires the next time an agent tries the same mistake.
- SkillEvaluator: NVIDIA's multi-tier framework for evaluating AI agent skillsMulti-tier grading of agent skills with quality gates — the skill score behind each PR verdict.
- t3code: one control surface for every coding agent on your machineOne control surface across Claude Code, Codex, Cursor, OpenCode and more — the provenance source for which agent wrote what.
- Pulsar: cross-harness agent-to-agent messaging for coding agentsCross-harness agent messaging so the gate can ask the authoring agent to fix a failing check instead of just blocking.
Repo to start from
agentgate — a GitHub Action and app that grades agent-authored PRs on provenance, skill quality, and repo incident history, and blocks the merge when the risk score clears the team's bar.
Evidence
- StarSling: self-improving CI for agent-volume repos
- SkillEvaluator: NVIDIA's framework for evaluating AI agent skills
- t3code: one control surface for every coding agent
- CodeRabbit pricing: Essentials $24/developer/month
- When an agent makes a mistake twice, we make it a check (Phoenix/Rush)
- AI coding agent mistakes: what broke and the lessons learned
- AI coding agents are writing more code. Why are senior engineers reviewing more of it?
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.