XYZ AI Lab Drops a Matched Pair of Long-Horizon Agent Repos on the Same Day — a Runtime and Its RL Trainer

GitHub·medium signalXBlueskyLinkedInCopy link

XYZ-AI-Lab published AxisAgentic (855 stars) and axrl (718 stars) both on 2026-07-23, and neither README mentions the other despite being two halves of one loop. AxisAgentic is a runtime that emits append-only traces reconstructing exactly what the model observed at any point, with explicit runtime markers for rollback, context compaction, and discard events — which is what makes state-faithful SFT export possible without leaking hidden history into the training example. axrl is the post-training side: SGLang rollout plus Megatron training, with PPO, GRPO/GRPO2, GSPO, TOPR and TIS objectives, black-box harness integration via an OpenAI-compatible proxy, and trajectories running 300+ turns.