Web-scale parallel onboarding and budgeted routing calibration for MAS

Develop a Web-scale, cost-budgeted protocol for continuously and in parallel onboarding and expanding the agent pool in large language model-based multi-agent systems, including mechanisms to warm up newly added agents and calibrate router decisions so that end-to-end performance does not degrade during expansion.

Background

The paper addresses performance collapse in LLM-based multi-agent systems when new agents are added without expansion-aware familiarization and conservative router updates. MonoScale proposes agent-conditioned warm-up tasks and trust-region memory updates to ensure monotonic non-decreasing performance during sequential expansion.

While experiments demonstrate stable scaling up to 10 agents, the authors note that realistic Agentic Web settings involve thousands to millions of agents with retrieval-driven routing. Scaling continuous onboarding and calibration to this Web-scale regime under a controllable cost budget, while preventing performance degradation, remains unresolved.