Agent Learning: Execution-Aware Policy Optimization for LLM Tool Systems

Formalizes tool orchestration in LLM systems as a learnable policy optimized for execution cost, latency, and constraint satisfaction.