Radar / Ideas / Agent-proof merge queue: ship PRs that…

Agent-proof merge queue: ship PRs that survived production

daily ideaambitiousJEV confidence 0.51JEV monetization 0.572026-10-15
OutcomeA main branch that stays green while agents merge 10x more PRs.

The problem

AI coding agents now produce most pull requests, and review quality is collapsing: tldraw auto-closes external PRs, QEMU suffered 125 AI reports in 10 minutes, and an ACM-tracked 2026 study found PR volume 6.8% above expected with falling merge rates. Senior engineers spend reviews deep-debugging plausible-but-wrong code instead of verifying it. Anthropic hit 30,000 agent merges a day and had to rebuild Git itself.

The idea

A merge queue where every batch is validated three ways before landing: (1) Decision 3.0-style typed gates answer 'safe to merge? score the risk' instead of parsing LLM prose; (2) a Polylane-style production lens checks the diff against live infra, dependencies and telemetry; (3) a Kerno-style runtime pass executes the app against the real stack and feeds failures back to the coding agent in-session. Scion-style server-side checkouts keep the queue fast at agent scale. The queue reports a merge-risk score with evidence per batch, and blocks high-risk ones.

Why now

Anthropic's Scion post (Oct 13) proves agent-scale merging broke Git and shows exactly what server-side infrastructure agents need. Decision 3.0 (Oct 10) makes typed, auditable merge gates cheap and open. Polylane (Oct 15) validated the production-grounded review concept. Kerno 1.0 (Oct 15) showed runtime execution catching bugs mocks miss. The components all shipped this week.

What it combines

Collision of four capabilities: Scion's server-side agent-scale checkouts (anthropic-scion-source-control-rebuilt-for-agent-scale-development) + Decision 3.0's typed decision models as auditable gates (decision-3-0-open-multimodal-decision-models-0-6b-27b) + Polylane's live-production evidence (polylane-impact-analysis-pr-review-against-live-production) + Kerno's runtime QA (kerno-1-0-ai-runtime-backend-qa-engineer). The mix matters because merge safety needs all four: fast checkouts, typed gates, production context, and real execution; each alone is a different product.

MVP

Weekend MVP: a GitHub App that queues PRs with Mergify-style rebase-and-test, then runs a Decision 3.0 judgment call over the diff (safe/risky + score) and posts an evidence comment. Deliberately skip: production-telemetry integration, runtime execution, server-side checkouts. The MVP de-risks the gate; the other three layers are the roadmap.

How it makes money

Budget owner: VP Engineering / Head of Platform, from the existing dev-tools budget. Pricing anchor: Mergify charges $21 per active contributor per month for its merge queue (https://mergify.com/product/merge-queue); Aviator's monorepo-aware queue is $50/dev/mo (https://www.aviator.co:443/aviator-mergequeue-mergify). Willingness-to-pay: 477 organizations already run Mergify queues across 200,000 merges (their own data, same page); tldraw closed all external PRs and QEMU was DDoSed by AI reports, showing teams pay to fix this pain.

The moat

GitHub, the largest incumbent, cannot ship this in one quarter: its merge queue validates against GitHub Actions only, while this product's core data advantage is live production telemetry and runtime execution across any CI provider, which GitHub structurally does not own and cannot bundle without a production-monitoring business it has never built. It would also cannibalize Actions compute minutes, a real revenue line, creating business-model conflict.

Distribution

B2B: wedge is teams whose agent PR volume already overwhelmed reviewers (the tldraw/QEMU class of teams). Sell per-repo to VP Engineering / platform teams via GitHub App marketplace, plus an agent-skill (like Polylane's CLI skill) so coding agents self-queue their PRs.

Why it wins

Existing merge queues (Mergify, Aviator, GitHub's native queue) only validate the combined diff against CI, which structurally cannot catch config drift, capacity assumptions, or runtime-only bugs. Code-review agents (CodeRabbit, Greptile) review statically. This queues on production evidence plus runtime execution, not CI green.

Risks

Teams may distrust automated gates after being burned by noisy AI reviewers; the MVP de-risks this by starting as advisory comments and only auto-blocking when a team's historical data shows precision. Also, production-telemetry access is a hard enterprise sell, so v1 targets repos with staging environments instead.

Build it with

Repo to start from

mergegate-live: a GitHub App + CLI that queues PRs through typed JEV-style risk gates with production evidence.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.