New Meta paper finds that having 2 different coding agents review each other's patches catches far more silent bugs than giving 1 agent a bigger budget.

Mixing Claude Code and Codex on the same task raised fully correct patches from 45.8% to 62.5% at matched spend.

Agents editing real training code can leak test data, break a gradient, or miswire a flag.

The code still runs, so you burn GPU hours and get numbers that look valid.

The reason is uncorrelated mistakes: agents from the same product fail the same way, so there is little for review to catch. The effect repeated on an unrelated training codebase.

– arxiv. org/abs/2609.39551

Title: "RankEvolve: A Reliable Multi-Agent Auto-Research Harness for Evolving Ranking Models"