Title: Mastering the Development of Functional Agentic Systems

Abstract:

While the automated design and optimization of LLM-based agentic systems can yield sophisticated models that significantly outperform standard off-the-shelf patterns, real-world deployments reveal a different priority. Fielded systems tend to emphasize simplicity, controllability, and the predictability of inference costs over raw performance. To address this gap, this paper introduces principled methodologies for designing and optimizing practical agentic systems.

We present an agent framework that allows designers to enforce modularity through the use of "pseudo-tools." These tools invoke LLMs recursively but operate within a restricted context. Leveraging this structure, we hand-engineered agents across a diverse array of tasks. Our findings indicate that, compared to dynamically planned workflows, these hand-constructed fixed workflows typically offer greater accuracy at a lower cost.

Furthermore, we introduce novel learning techniques tailored to the specific agentic components required by our framework, including pseudo-tools and fixed workflows. These learning-driven approaches generally surpass the performance of hand-engineered agents. Finally, we utilize the framework’s inherent modularity to implement multi-objective optimization, allowing for the simultaneous tuning of cost and response quality, as well as the integration of outputs from multiple learning systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC