Governance as a Safety Net : An LLM Agent-Based Study of Open-Source Community Sustainability

Open-source software (OSS) sustainability depends not only on technical capability but also on how communities allocate work, distribute authority, and retain contributors. Existing agent-based models of OSS communities typically abstract developer behavior through fixed rules, while recent LLM-based multi-agent systems provide richer social interaction but often lack persistent grounding in executable software environments. We introduce LLAMOSC, an LLM-powered agent-based framework that couples contributor-level psychological state, governance mechanisms, social interaction, and autonomous code execution within a reproducible simulation environment. LLAMOSC evaluates two contrasting governance topologies: centralized maintainer-assigned task allocation and decentralized peer-bidding. Agents operate on software-development tasks through a grounded AutoCodeRover execution layer, receive automated contribution-quality feedback, and update internal motivation states in response to allocation, execution, and review outcomes. Across 60 simulation runs spanning three development environments, decentralized governance exhibits its strongest motivational effect in the lower tail of the contributor distribution rather than through a uniform increase in motivation. Severe simulated burnout, operationalized as final motivation $M_i < 3.0$, decreases from 31.3% under maintainer-assigned governance to 18.7% under decentralized peer-bidding ($p = 0.0267$). Decentralized governance is also associated with more distributed workload allocation and stronger performance as task difficulty increases. These findings suggest that governance topology can function as a socio-technical feedback mechanism: decentralized task allocation does not eliminate technical difficulty, but may provide a "safety-net" effect that reduces extreme contributor burnout depletion while redistributing coordination across the community.