Parallel Coding Agents Break Each Other in 97% of Constructed Django Tasks, and One Message About the Concurrent Change Recovers 82%

arXiv·medium signalXBlueskyLinkedInCopy link

The 'stale' benchmark runs tests on each agent patch alone and then merged, counting only failures the combination introduces. On constructed tasks using 12 real Django helpers, interference occurred in 97% of runs, and a message describing the completed concurrent change recovered 82% of them. Among 834 runs on 417 mined real Django PR pairs, only one showed interference, so the authors caution the constructed rates are not a real-world frequency. The practical fix for multi-agent coding setups is to broadcast interface changes between workers.