BUILDING DEEP RESEARCH: HOW WE ACHIEVED STATE OF THE ART

Summary from the TLDR Dev newsletter

Tavily rebuilt its research agent from scratch after overengineering the first version with assumptions that broke when new models arrived. Instead of passing all tool outputs through the agent loop like most systems do, it distilled outputs into compact reflections and only loaded raw sources for final generation. This cut token usage by 66% while hitting SOTA on benchmarks.