Deterministic rule-guided dispatch beats autonomous code-review agents by 2.17x SEM-F1 on 5–15x fewer tokens

arXiv 2608.09290·high signalXBlueskyLinkedInCopy link

OpenCodeReview (arXiv 2608.09290, Aug 10) argues LLM code reviewers fail on two axes — non-determinism from unbounded tool use, and context locality that caps issue depth at the diff — and fixes both by injecting structure at three pipeline points rather than maximizing autonomy. A multi-layer rule system deterministically selects which files get reviewed and against which criteria, grounded per-file review runs on a curated toolset, parallel subagents recover cross-file dependencies, and an Independent Reflection stage applies a falsification-first filter under an asymmetric information boundary to strip hallucinated comments. On 200 real pull requests across 10 languages with 1,505 expert-verified comments it hit 25.10% SEM-F1 versus 11.57% for baseline agents while consuming 5–15x fewer tokens — the pattern generalizes: put the triage in rules, spend the model budget on grounded reading.