Stop running everything through LLMs: How I fixed my research agent

TinyResearch is presented as a research tool for people and agents that need dependable answers to complex questions, rather than a language model asked to handle search, reading, and synthesis all at once. Its design is judged against factual correctness, source credibility, readability, comprehensiveness, and instruction following. For a question such as whether to buy a Tesla Model 3 or Toyota Camry, it breaks the request into smaller questions, searches for evidence, and checks whether important topics - such as long-term reliability or charging access - remain unanswered before writing a report.

The research loop plans queries, finds and ranks candidate sources according to the specific question, reads pages live, and extracts claims linked to their sources. It revisits gaps and builds visible partial answers along the way; the final report uses only collected claims and includes numbered citations. Domain controls can prioritize, exclude, or restrict research to specified sites, with filters applied throughout the process. A typical deep run works through about eight questions per round, ranks roughly 100 sources per round, performs around 40 searches, and reads about 80 pages, usually in under 15 minutes. Runs have time and step limits, while MAX mode allows deeper investigation. One architectural change was to remove a separate planning component: handing plans between separate systems lost context, especially when users narrowed a follow-up request, so planning moved into the orchestrator.