Why green CI can still break main when AI agents open parallel PRs

CI can be green on two pull requests and still leave main worse after both land. That sounds contradictory only if we treat each PR as an isolated unit.

Modern AI coding agents make the gap easier to hit. One agent changes a route, another updates the code that uses it, and a third touches a nearby setup file. Each PR may pass its own tests. The real risk is the relationship between open PRs and the order in which they land.

What CI sees

CI usually answers one question: does this branch pass when tested against the base it knows about?

That is necessary, but it does not always answer:

• which other open PRs are working in the same area;
• whether one PR should land before another;
• whether the safe order is A then B, B then A, or wait and revise;
• whether an author saw the warning before choosing to merge.

This is why a small team can feel surprised by main breakage even when every PR looked reasonable in isolation.

What Veripsa adds

Veripsa Core watches the space between open GitHub pull requests. It posts a GitHub check/comment when a PR may need another in-flight PR to land first.

The important part is not a dramatic claim that every failure disappears. The useful part is a small control point before merge:

• Clear means Veripsa did not find a covered reason to slow down.
• Unknown means the signal was not strong enough to call clear.
• Wait in line means another open PR appears to be in the same landing lane.
• An optional veripsa-ack label can record a deliberate override when a team wants an audit trail; normal advisory use does not require it.

By default, Veripsa is advisory. Teams can choose whether branch protection treats that check as a hard gate.

Why this matters for AI-agent work

Parallel AI-agent work increases throughput, but it also increases the number of active branches touching nearby product surfaces. The bottleneck moves from "can one PR compile?" to "can multiple reasonable PRs land in a safe order?"

That is the point of Veripsa: pre-merge traffic control for the gaps between PRs, without becoming a code reviewer or storing customer source bodies.

For a deeper Japanese article on the same problem, read GitHub Actionsが緑でもmainが壊れる理由.

Start here:

• How Veripsa works
• AI-agent PR collision
• Working with agents