Learning to Construct Practical Agentic Systems

Researchers propose a practical framework for building LLM-based agentic systems that prioritizes simplicity, cost predictability, and controllability over maximum optimization. The framework uses modular "pseudo-tools" and fixed workflows, demonstrating that hand-engineered agents often outperform dynamically-planned systems in production environments.

This research addresses a critical gap between academic optimization of LLM agents and real-world deployment constraints. While recent literature has focused on sophisticated automated design and dynamic planning of agentic workflows, production systems reveal different priorities: cost predictability, reliability, and maintainability trump marginal quality improvements. The authors argue that modularity through pseudo-tools—LLM calls restricted to specific contexts—enables engineers to build agents that are easier to understand, debug, and control in practice.

The framework's emphasis on hand-constructed fixed workflows over dynamic planning reflects maturation in the AI systems space. In emerging technologies, initial academic focus typically gravitates toward maximum capability and optimization. However, as systems move to production, stakeholder concerns shift to operational concerns. This pattern mirrors early adoption curves across computing domains.

The research demonstrates that learning methods applied within their modular framework outperform hand-engineered baselines, suggesting a middle path: structured systems that remain interpretable while benefiting from optimization techniques. The multi-objective optimization approach—jointly optimizing cost and quality—directly addresses enterprise deployment realities where inference costs matter as much as accuracy.

For developers and AI teams, this research validates a pragmatic approach to agentic systems: start with simple, fixed workflows, apply learning methods within modular boundaries, and explicitly trade off quality against cost. Organizations building production AI systems should expect that sophisticated dynamic planning adds complexity without proportional value gains, making the case for structured, optimizable architectures.

• →Hand-engineered fixed workflows often outperform dynamically-planned agentic systems in production, suggesting simpler architectures are competitive for real-world deployment.
• →Modularity through pseudo-tools enables cost predictability and controllability, addressing the gap between academic optimization and production system requirements.
• →Learning methods applied within structured frameworks generally outperform hand-constructed agents, providing a path for systematic improvement without sacrificing interpretability.
• →Multi-objective optimization jointly addressing inference cost and response quality reflects how enterprise AI systems must balance performance against operational budgets.
• →The research suggests the AI systems field is maturing from pursuing maximum capability toward pragmatic design patterns that prioritize maintainability and reliability.