Million-Agent Ruby: Ractor-Local GC in the Age of AI

EN

Ractors promised true parallelism, but "Stop-the-World" Garbage Collection remained the final bottleneck—until now. Ruby 4.x introduces Ractor-Local GC, a monumental shift in memory management that decouples object heaps and allows for disjoint collection. This talk explores the technical breakthroughs of the MaNy Project (M:N threading) and how they enable "Hyper-Generational" on device workloads in Ruby. Using ragents—a Ractor-native AI agent framework—we demonstrate how to bridge the conceptual gap between Memory Management and LLM Context Management. We will visualize the synergy between heap compaction and token window summarization, proving how Ruby 4.x enables near-linear scaling for autonomous agent swarms. Attendees will leave with a deep understanding of Ruby 4.x internals and a blueprint for high-concurrency AI orchestration.

  • Justin Bowen
    Justin Bowen

    Rubyist and AI Researcher with over a decade of experience focused on high-concurrency systems and GPU accelerated computing in Ruby. Pushing for AI tooling in Ruby to compete with Python. I am the creator of Active Agent, Solid Agent and now ragents, a Ractor-native framework for autonomous AI agents. My work focuses on bridging the gap between Ruby's VM internals and the practical demands of modern AI inference and orchestration, specifically regarding memory locality and parallel execution.