Want to train your own Claude Code/Codex agent with your own model? We are excited to roll out ProRL Agent V2: Polar.
An infrastructure for black-box agentic RL, Polar lets you train agents with any harness, whether it’s OpenClaw, Hermes, or a custom agent built with frameworks like LangChain, Autogen, AG2 and others.
Check out here:
Code: https://t.co/AdfJHWE1Gp
Paper: https://t.co/5SnzJX2wCZ
Welcome to the world of agentic RL, without opening the box.
Inspired heavily by NVIDIA’s POLAR work, we just shipped our POLAR-style rollout pipeline for Subnet 66.
This is a big step toward turning live coding-agent competition into training data.
NVIDIA’s POLAR paper made the point really well: the useful data is not just whether an agent got the answer right. It is how the agent got there: the steps it took, the tools it used, the mistakes it made, and the way it recovered.
Our pipeline now records solver trajectories from live validator/task-pool runs, links them to task results, redacts them for public release, and exports retired rollout bundles to Hugging Face.
The goal is to preserve real agent behavior from competitive environments instead of relying only on synthetic offline traces. Better rollout data should make it easier to iterate on agents and improve them faster.
The published rollouts will be useful for anyone working on improving coding agents, not just Ninja miners.
We just shipped xninja v0.1.0.
Think Claude Code, but powered Subnet 66's top agent harness. Open a project, ask for help, let it inspect the code, make changes, and hand you a patch to review.
It’s one step of many toward bringing Subnet 66 agents out of benchmarks and into real developer workflows.
Try it out with
`pip install xninja`
Miners on Ninja are building a coding harness grounded in real-world engineering usage.
By scraping GitHub commits, issues, PRs, and repo activity, Ninja generates tasks from real software development. Miners improve the harness by making it better at solving these real coding problems, and every winning improvement compounds into the next version others build on.
The goal is to use open competition and real-world problems to create the leading open-source harness for coding.