Idea: Prompt "gradient" based on workflow past patterns

[Fosowl]

We can extract meaningful pattern from the workflow library code and state_results such as mean KL divergence between agent prompt, lengh, complexity, type of agents, etc and how it relate to the score/performance. Storing these pattern after workflow execution would allow to get information such as :

Correlations with Final Reward

Feature	Pearson r	Spearman ρ	p-value	Interpretation
num_agents	-0.158	-0.117	0.0007***	More agents = worse
mean_js	-0.148	-0.129	0.0015**	Lower JS divergence = better
entropy_gradient	-0.097	-0.080	0.0393*	Negative gradient = better
max_kl	-0.104	-0.094	0.0267*	Avoid sharp peaks
mean_kl	-0.095	-0.087	0.0424*	Lower divergence = better
mean_cosine	+0.094	+0.089	0.0446*	Higher similarity = better

With value thresholds this can be used to seed the workflow generation prompt with tips such as:

• Increate prompt lengh
• Ideally use 3 agents
• Reduce prompt complexity
• Use more similar agent

Providing search direction for generating the next workflow variant