Declarative Skills for AI Agents in Knowledge-Grounded Tool-Use Workflows

In three linesComparative study of three orchestration paradigms for tool-using AI agents in customer-service workflows: declarative agents (natural-language skill files), imperative agents (state machines), and unscaffolded baseline. Tested on 5 LLMs and 2 retrieval regimes. Finding: retrieval quality is the dominant bottleneck; declarative skills consistently improve accuracy on procedural tasks when retrieval is high-quality.

Summary generated by Claude — human-verified