Google Research and Northwestern build self-improving AI system Harness²

Google Research and Northwestern University introduced Harness², a framework that lets AI agents refine their own workflows and tools without additional training. Using a Propose–Probe–Compose loop, modular skills, custom tools and sub-agents, smaller models can outperform larger ones on specialized tasks.

• Propose–Probe–Compose loop refines the harness after each task
• Modular skills and custom tools activate only when needed
• Sub-agents manage large datasets without context loss
• Limitations include high compute demands and autonomy risks