Coding agents raise open-source throughput 39% but collapse human-to-human interaction from 32.4% to 11.6%

arXiv·medium signalXBlueskyLinkedInCopy link

An LLM-based multi-agent simulation seeded with real GitHub data from 1,084 active developers branched the same community state into parallel no-agent and agent conditions for 4-week runs. Coding agents raised planned tasks 34.0% and completed tasks 39.0% and cut median completion time from 45 to 20 minutes, but adoption reached only 26.0% and gains concentrated among already-active, well-connected developers. The structural cost is the finding: direct human-human interaction fell from 32.4% to 11.6% while agent-mediated modes rose to 57.3% (40.3% of it agent-assisted self-loops), and the resulting public corpus scored 22.3% knowledge coverage on a retrieval benchmark versus 81.1% for the real-human corpus.