HarborRL update adds elastic scaling, SSH-free cluster deployment for agentic RL training
Dear_Treat3688 · reddit · 2026-11-14
HarborRL, an open-source (MIT) agentic RL framework from datamllab, shipped a November 2026 update focused on simplifying distributed training and deployment:
- Elastic scaling: gateway scaling with static routing, automatic cross-machine scaling, and auto-retirement of idle resources
- SSH-free cluster deployment: a Hub mode where workers pull tasks over HTTP, working even behind firewalls
- Unified RL pipeline: multiple RL algorithms integrated into a shared Slime-based pipeline, validated on math reasoning tasks
- Better config & sandboxing: improved setup wizard, web console, sandbox rollout, and resource cleanup
The framework targets the heavy engineering of connecting task environments, agent harnesses, and RL training engines in distributed settings. Feedback requested on rollout execution and sandbox management practices.
More from coding & agent
- FastMCP 4.2 ships Skills as first-class MCP objects, delivering expertise alongside tools — EdenEmarco177 · 2026-11-14
- Dev builds MCP bridge linking web AI chats to local code editors, seeks security feedback — Quiet-Baker8432 · 2026-11-14
- ThePrimeagen visualizes program state machines to regain control in the agent era — samgoodwin89 · 2026-11-14
- distilled goes self-hosting: tool now fetches AWS schemas via its own generated package — samgoodwin89 · 2026-11-14
- korea-data-mcp Resolves Korean Company Identity Traps: 13,530 Firms Share 5,573 Names — youngtipted · 2026-11-14
- 10 open-source GitHub repos to deploy AI agents with HTTPS, retries and Kubernetes — victor_explore · 2026-11-14