XYZ AI Lab ships runtime and RL trainer same day

XYZ AI Lab shipped two open-source agent tools on July 23: a runtime and a reinforcement-learning trainer, per their GitHub repos. Together they cover the loop needed to train long-horizon agents: record what a model did, then train on it. AxisAgentic already has 855 stars and axrl 718, a sign teams building these agents needed exactly this.

AxisAgentic logs append-only traces of everything an agent model observed, with explicit markers for rollbacks, context compaction and discarded turns. That makes it possible to export clean fine-tuning data from a live run without leaking the hidden history those markers strip out.

axrl is the training half, SGLang for rollouts, Megatron for the actual training. It supports six objectives, PPO, GRPO, GRPO2, GSPO, TOPR and TIS, with trajectories running past 300 turns. A black-box harness can plug in through an OpenAI-compatible proxy.

Neither README mentions the other, per the repos. Yes, but axrl doesn't require AxisAgentic specifically. Its proxy takes rollouts from anything that speaks the OpenAI API, agent runtime or not.