How they use Tavily
Forecasting Research Institute uses Tavily web search and page reading in its FRI Forecaster package. AIRO and the ForecastBench driver use this package for agentic forecasting. AIRO forecasts AI-related catastrophic-risk questions on a twice-weekly schedule. Its runner requires a Tavily key for grounded forecasts and refuses to run without one unless ungrounded answers are explicitly allowed.
FRI's AIRO announcement describes models choosing their own follow-up searches and allowing forecast submission after at least 10 research-tool calls have returned. AIRO enables controls that withhold submission tools until research-floor and search-coverage requirements are met. Since September 2, 2026, each model is asked once per run.
The forecasting loop records research evidence alongside answers and usage. Its Tavily tools cache searches and full-page extracts for one hour by default, sharing identical requests across local workers. The package now lives in the AIRO repository after moving from the archived agent-box repository.
Non-commercial use. FRI uses Tavily search and page reading for nonprofit forecasting research.
Sources
4 sourcesAIRO forecasting implementation
Listed 31 Oct 2026 · GitHub · Web page
AIRO requires Tavily for grounded forecasting runs. Since September 2, 2026, it asks each model once per run and redirects the budget previously used for repeated same-day elicitations into its agentic search harness.
Open original sourceAutomating Catastrophic Risk Forecasts: Introducing AIRO
Published 11 Sept 2026 · forecastingresearch.substack.com · Web page
FRI says AIRO models independently research catastrophic-risk questions using Tavily web-search and page-reading tools. Each model chooses follow-up searches and can submit forecasts after at least 10 research-tool calls have returned.
Open original sourceFRI Forecaster original repository
Listed 31 Oct 2026 · GitHub · Web page
Documents Tavily search and page-reading tools used by FRI's forecasting package for AIRO and the ForecastBench driver. The repository records the package's move into AIRO.
Open original sourceCompany context
forecastingresearch.orgForecasting Research Institute is a nonprofit research organization that develops and tests forecasting methods for decisions involving AI, biosecurity, and geopolitics. It runs expert forecasting studies, builds automated forecasting tools, and evaluates the accuracy of AI predictions. Its projects include ForecastBench, an AI forecasting benchmark; the Longitudinal Expert AI Panel (LEAP), which tracks experts’ expectations about AI progress and impacts; and the Automated AI Risk Outlook dashboard. Its research informs policymakers, researchers, and the public.
Recent updates
2 updatesFRI launches AIRO catastrophic-risk forecasting dashboard
FRI launched Automated AI Risk Outlook, a dashboard that uses an ensemble of frontier AI models to forecast catastrophic risks. It publishes individual model forecasts and ensemble estimates across multiple time horizons.
Forecasting Research InstituteFRI publishes Anthropic-funded biosecurity safeguards study
FRI released research completed for Anthropic assessing its ASL-3 safeguards. The study surveyed 22 domain experts and 20 superforecasters about potential reductions in biosecurity risks.
Forecasting Research Institute