Training a harness for model-agnostic and task-environment-agnostic capability improvements with PyTorch-like framework [P]

Executive Summary

I worked on this project (https://github.com/workofart/harness-training) for the past few months to reframe "Agent-driven Self-improving Harness" to "Harness Training". The idea is simple, the harness is trained once with a frozen task LLM against a given task environment. Then you can then swap out...