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

Researchers compare three orchestration approaches for AI agents handling customer-service workflows: declarative agents using natural-language skill files, imperative agents with programmatic state machines, and unscaffolded baseline agents. The study finds that retrieval quality is the dominant bottleneck, and declarative skills improve performance on procedural tasks only when evidence quality is high.

This research addresses a critical challenge in deploying AI agents for real-world applications: how to structure their decision-making processes effectively. The study compares three fundamentally different architectural approaches, revealing that the method of orchestration matters less than the quality of underlying information retrieval systems.

The declarative approach—using natural-language skill files—represents a more flexible paradigm than traditional programmatic state machines. This aligns with broader trends in AI development favoring interpretable, modular systems that can be adjusted without code changes. The research formalizes these architectures within a decentralized partially-observable Markov decision process framework, providing theoretical grounding for practical design choices.

For developers and organizations building AI-powered customer service systems, the findings carry important implications. The research demonstrates that optimizing retrieval mechanisms should be the priority before implementing sophisticated orchestration strategies. When retrieval systems fail to surface relevant information, even well-designed skill files cannot compensate. Conversely, with high-quality retrieval, declarative agents consistently outperform brittle imperative approaches, suggesting that modularity and flexibility yield tangible benefits.

The brittleness of imperative state machines presents a practical warning: rigid control flows may seem appealing initially but create maintenance burdens and failure modes as systems scale. Looking forward, the field should focus on improving retrieval quality as the foundation for agent performance, while adopting declarative approaches that allow graceful degradation and easier iteration. Future work should explore hybrid architectures combining the strengths of both paradigms.

• →Retrieval quality is the dominant performance bottleneck for AI agents, limiting the effectiveness of orchestration strategies alone.
• →Declarative agents with natural-language skill files outperform rigid imperative state machines under high-quality retrieval conditions.
• →Skill files cannot recover lost performance when underlying evidence is incomplete or skewed.
• →Declarative orchestration reduces errors on procedural tasks and provides more flexibility than programmatic approaches.
• →Future AI agent development should prioritize retrieval system optimization before implementing complex control flow architectures.