Yukta Kulkarni is a researcher known for contributions to multi-agent systems and large language model orchestration research. Kulkarni has collaborated on empirical studies examining how different architectural patterns impact the performance and efficiency of AI agent systems.

Kulkarni's research concentrates on the practical evaluation of multi-agent system architectures, with particular emphasis on orchestration patterns that coordinate behavior across multiple language model instances. This work addresses a critical challenge in AI systems: determining optimal coordination strategies when deploying agents to handle complex, real-world tasks at scale.

In collaboration with co-author Siddhant Kulkarni, Kulkarni contributed to a comprehensive benchmark study evaluating agent orchestration approaches. The research examined four distinct orchestration patterns (architectural approaches for coordinating multiple agents) across a substantial dataset of 10,000 SEC filings — regulatory documents from publicly traded companies that present complex information extraction and analysis challenges ¹⁾.

The empirical evaluation tested these patterns across five different large language models, providing comparative analysis of how different LLMs respond to various coordination strategies. SEC filings represent a particularly demanding use case for multi-agent systems, as they require parsing dense financial and legal information, cross-referencing multiple document sections, and synthesizing insights across unstructured text at scale.

Research into orchestration patterns addresses practical deployment challenges for AI systems. As organizations scale AI implementation beyond single-agent applications, understanding which coordination patterns optimize for accuracy, latency, cost-efficiency, and robustness becomes increasingly important. The use of real-world SEC filing data grounds the research in authentic use cases rather than synthetic benchmarks, providing insights applicable to document processing, financial analysis, and information extraction applications.

The comparative evaluation across multiple LLM implementations helps practitioners understand whether orchestration effectiveness depends on specific model architectures or whether certain patterns demonstrate consistent advantages regardless of underlying model choice. This distinction has significant implications for system design, vendor selection, and resource allocation in production environments.

• Siddhant Kulkarni
• arXiv:2604.18071
• Token Efficiency Across Orchestration Patterns
• Multi-Agent Orchestration
• Deepagents