Signal From Social Coding to Agentic Coding: Productivity and Relational Reconfiguration in Open-Source Communities

Summary

Mengying Zhou, Yongjie Yin, and Yang Chen published a paper on arXiv (cs.CY) on August 4, 2026, studying how generative coding agents (CAs) reshape open-source software communities using an LLM-based multi-agent simulation initialized with real GitHub data from 1,084 active developers. After a warm-up period using historical commits, the researchers branched the same community state into parallel conditions with and without coding agents for four-week simulations. CA introduction increased planned and completed tasks by 34.0% and 39.0% respectively, and cut median task completion time from 45 to 20 minutes, but adoption reached only 26.0%, with gains concentrated among developers who were already more active and well connected. CAs also restructured task execution pathways: direct human-to-human interaction declined from 32.4% to 11.6% of interactions, while CA-involved modes rose to 57.3%, including 40.3% completed through CA-assisted self-loops. Critically, public knowledge generated under the CA condition provided less support for later tasks, with the CA corpus achieving only 22.3% knowledge coverage on a standardized retrieval benchmark versus 81.1% for the real-human corpus, requiring more retrieval steps and achieving lower success rates. The authors describe this as a 'productivity-public knowledge tension' in which coding agents boost technical output while shifting more work into agent-mediated or private loops that leave less useful public records for future contributors.