Why keep the matcher's output dull?

Because the reviewer treats anything in the matcher's file as a conclusion already reached. Atlassian's rule is file paths and guideline names, nothing else. The second file follows the same rule: a list of what was excluded, with no opinion attached, so the humans reading it draw their own.

What goes in the reject pile for a review pipeline?

Three columns. Changed files the matcher mapped to no guideline, which is where a missed comment lives. Guideline files not routed to in the last 30 days, which are covering a dead directory or being skipped. And pull requests where the matcher returned nothing in scope at all, counted. The owning team reads it on the same weekly cadence they read the comments.

Why not skip the matcher when there is only one guide?

Atlassian suggests skipping it, since the matcher would pick the same document every time. With one guide the matcher is still deciding which changed files are in scope, and the out-of-scope list is the only record of what the reviewer never saw. One extra file is cheaper than a pipeline that gets quieter as it gets worse.

What a Two-Agent Code Review Pipeline Misses

What does a two-agent code review pipeline miss?

THE SHORT ANSWER

The comment that was never written. In Atlassian's Agentic Pipelines setup, per Arthy Vijayaraghavan, a matcher pairs each changed file with the guideline that covers it and a reviewer opens only what the matcher names. The feedback loop edits a guideline whenever a comment is wrong, but a change routed to the wrong document or to none produces no comment, so nothing gets corrected. The fix is a second dull file beside the first: changed files mapped to no guideline, guidelines not routed to in 30 days, and pull requests where the matcher returned nothing, read weekly. My own reading pipeline has the same shape; it read 221 posts this week and dropped 136, and every bug I have found in it was in the drops, not the keeps.

A two-agent split is the right fix for an agent that holds every guideline at once and comments on code nobody touched. The hand-off creates a new blind spot, though: the reviewer can only be wrong about what it was shown, and the loop that fixes guidelines only fires on comments that exist.

The counts from my own pipeline, the three bugs that only the drops revealed, and the three columns of the second file are in Atlassian Split the Reviewer in Two. Ship the Matcher's Reject Pile. The rule it rests on is A Name on the Definition, and the Reject Pile.

Last reviewed 2026-10-15 · 1 min read

What does a two-agent code review pipeline miss?

SOURCES

THE LONG VERSION

RELATED ANSWERS