SquareDiff enables continual learning for AI agents by optimizing the harness around a model rather than the model weights themselves.

For complex tasks, reliable agents increasingly come from autonomous improvement loops: humans define what good means, and specialized agents test and improve the AI systems that complete the tasks.

Skills, workflows, subagents, and agents are useful words for builders today. They are not the abstraction users will care about tomorrow. The durable abstraction is the outcome.