TinyFish publishes agentic search architecture guide and deep research lessons

Across seven QA datasets the multi‑turn search component beat RAG baselines by 41% with Qwen2.5‑7B and 20% with Qwen2.5‑3B. In a search‑only benchmark TinyFish completed 59.3% of tasks, compared with 77.3% for Perplexity, 70.3% for Firecrawl and 66.3% for Exa Fast. Engineering changes such as merging planning into the orchestrator, removing sub‑agents, and ranking sources by research objective reduced latency by up to 5x and increased throughput by up to 3.4x. The guides also outline failure‑aware retries, evidence contracts, and strict caps on tokens, time and cost.