Research agents are becoming essential in AI applications, transforming how we gather and synthesize information. Traditional human research is limited by memory and time constraints, while AI can process and analyze vast amounts of data quickly. The article outlines the lessons learned in developing a state-of-the-art research agent, emphasizing the need for a robust software layer to manage context, orchestration, and error handling as models evolve. The authors faced challenges with their initial architecture, which became a bottleneck as new models emerged, forcing a complete rebuild. The piece highlights advancements in model capabilities over the past seven months, particularly in tool-calling abilities. Future models are expected to address current pain points for developers, focusing on high-recall summarization and reliable tool interactions. Tools themselves should enhance interaction with large language models (LLMs) by providing relevant data while minimizing unnecessary context. Tavily has invested in an advanced search feature that improves the efficiency of data retrieval, reducing errors and latency. A key concept discussed is context engineering, which is vital for maintaining optimized context over long research tasks. Tavily's Advanced Search helps curate relevant content, ensuring agents do not become fixated on a single research thread. This approach includes global state persistence and deduplication of sources, allowing agents to access fresh information and broaden their research scope. The article argues for a more human-like iterative research process, where agents distill information and only return to original sources when compiling final outputs. The authors also address the challenges of building production-grade agents, balancing autonomy and performance with constraints like latency and reliability. They advocate for designing with non-determinism in mind, treating potential failures as integral to the process rather than as afterthoughts. They’ve managed to reduce token consumption by 66% compared to existing models while maintaining high performance on benchmarks, demonstrating a successful intersection of efficiency and quality.