PublicAgent: Multi-agent design principles from an LLM-based open data analysis framework

Open data repositories hold potential for evidence-based decision-making, yet remain inaccessible to non-experts who lack expertise in dataset discovery, schema mapping, and statistical analysis. While large language models show promise for individual tasks, end-to-end analytical workflows expose fundamental limitations: attention dilutes across growing contexts, specialized reasoning patterns interfere, and errors propagate without detection. We present PublicAgent, a multi-agent framework that addresses these limitations through targeted decomposition into specialized agents for intent clarification, dataset discovery, analysis, and reporting. This architecture maintains focused attention within agent contexts and enables validation at each pipeline stage. Benchmark evaluation across eight models and ablations across five representative models on 50 queries identify empirical design guidelines for multi-agent LLM systems in open-data analysis. The results show that specialization improves performance across the evaluated models, that discovery and analysis agents provide the most consistent value, and that report and intent agents show stronger model-dependent effects. Ablations further indicate that different agents mitigate distinct failure modes, with discovery and analysis failures often causing unusable outputs while report and intent removal more often degrades quality. These findings provide evidence for when and why specialization helps complex analytical workflows, while also highlighting the need for broader benchmarks, human validation, and continued evaluation as frontier models evolve.