A Quarkus app reviews a 2,900-line pull request for $0.45 by mixing plain code with agent debates and votes
Mario Fusco described on the Quarkus blog on 4 November how the LangChain4j team built an agentic reviewer for its own pull requests. The project had received 1,698 pull requests this year at the time of writing. The reviewer, a Quarkus application using the Quarkus LangChain4j extensions 1.15 on LangChain4j 1.21, reviewed a 2,900-line pull request for $0.45. It only reads GitHub with a read-only token and writes a Markdown report, the same report as JSON and an HTML trace of the agents.
The design choice that runs through the post is that the backbone is not a supervisor model. The root is a plain sequence of about twenty steps declared as a Java interface, and the extension checks at build time that every input an agent reads is produced by another step. A wrong key fails the build instead of a review. Many steps use no model at all: fetching the pull request, parsing the diff, validating findings and rendering the report are ordinary static methods.
Model judgment sits inside specific steps. A decision model, which returns probabilities for typed questions instead of text, picks which of five reviewers to run: correctness, security, design, tests, and documentation and API. With topic-style descriptions a 0.8-billion-parameter local decision model got 13 of 25 routing decisions right; after the descriptions were rewritten around simple flags such as whether production code changed, it got 22, and a hosted model got all 25. A deterministic safety net still runs a fallback correctness, security or documentation review when the router skipped one that the change requires.
Findings must point to a file in the diff and a line near a changed hunk, otherwise code discards them, which removes invented locations and text smuggled in through the pull request. Contested high-severity findings go to a debate between a prosecutor and a defender, with a judge deciding. Three scorers vote only when the first score card is borderline, and a non-AI calibrator can only make the outcome stricter. The report text goes through a critic and rewrite loop, while findings, scores and the verdict are rendered from data so the writer cannot invent or lose one.

What it means
The post is a practical argument for putting structure in code and judgment in narrow, checkable places. The cost figure comes from limiting the expensive patterns, voting and debate, to the cases that need them.