TinyFish reveals framework for reliable AI agent search systems

TinyFish, a provider of web infrastructure for AI agents, has released two technical guides detailing how to build production-ready agentic search systems. The guides focus on addressing common issues such as incomplete retrieval, failed retries, and execution limits. The architecture, built on four web primitives, operates through a seven-phase control loop around a base model. On OpenBenchmarks, TinyFish achieved a 79.0% completion rate on coding tasks with a median cost of $0.050, highlighting the importance of rigid control loops and unified reasoning. The system’s architecture draws from a February 2026 preprint that decouples reasoning from execution through typed tool interfaces. TinyFish splits a run into seven phases, with Phase 2 addressing the 'lost-in-retrieval' problem and Phase 6 leveraging Search-R1 to generate additional search queries. Across seven QA datasets, its learned multi-turn search outperformed RAG baselines by 41% and 20%. TinyFish also optimized its research system for quality and relevance, adjusting source ranking based on research objectives and adding controls for prioritizing, excluding, or restricting domains. The company developed an internal evaluation harness using an LLM judge to assess reports for relevance, evidence strength, and source diversity. OpenBenchmarks, updated on 26 Sept 2026, reports TinyFish’s search-only results at 59.3%, compared to 77.3% for Perplexity, 70.3% for Firecrawl, and 66.3% for Exa Fast.