Describe a behavior in words, and an agent writes the reward, trains a policy, judges the result via LLM-written code metrics and VLM, and revises until the policy matches your intent. No human intervention required.

The story of building a SigLIP2-class vision encoder from scratch using only open data.

Introducing Terminus-KIRA — boosting frontier model performance on terminal-bench with minimal harness improvements. We share our full methodology, failure analysis, and open-source release.

We propose an offline alternative using teacher-generated trajectories and a novel GRPO variant that better captures high-quality reasoning traces.