A step-by-step guide to building a multi-agent AI research assistant using the OpenAI Agents SDK, GPT-4o mini, and the Olostep Web API. The system uses a manager agent that orchestrates a judge agent (for evidence quality control) and an analyst agent (for report writing), with tools for web search and scraping. The guide covers environment setup, defining structured Pydantic output models, creating Olostep tool wrappers with tracing, wiring together the agent workflow, and exposing the finished assistant as a Reflex web app with PDF export. The architecture is designed to balance speed and cost by escalating research depth only when the judge agent deems initial evidence insufficient.
Table of contents
Introduction1. Set Up the Environment2. Test Olostep Search with Scraping3. Add Helper Functions4. Define Structured Output Models5. Create Olostep Tool Functions6. Build the Specialist Agents7. Create the Manager Agent8. Run the Research Assistant with Tracing9. Test the Multi-Agent Research Assistant10. Build a Web UI with ReflexFinal ThoughtsSort: