Multi-Agent Research Assistant
Autonomous research pipeline using specialized agents for search, synthesis, and citation
Overview
An autonomous research pipeline where a Planner Agent decomposes a research question into sub-queries, a Search Agent fetches real-time web results via Tavily, a Synthesis Agent extracts and merges relevant information, and a Writer Agent produces a structured report with citations.
Problem
Deep research on any topic requires multiple iterations of searching, reading, cross-referencing, and synthesizing — work that's time-consuming for humans and poorly handled by single LLM calls that lack real-time web access.
Solution
A four-agent pipeline with a shared research context object: Planner decomposes the topic → Search fetches live sources → Synthesis extracts claims and evidence → Writer assembles a structured, cited report. Each agent receives only the context slice it needs.
Architecture
- 1Planner Agent decomposes the research question into 3-5 targeted sub-queries
- 2Search Agent calls Tavily API for each sub-query, retrieving ranked, real-time web results
- 3Synthesis Agent reads all search results and extracts key claims, evidence, and source URLs
- 4Writer Agent structures extracted knowledge into a sectioned report with inline citations
- 5Shared ResearchContext object typed in TypeScript, validated at each agent handoff
- 6Orchestrator manages agent sequence and handles retries on structured output validation failures
Technical Challenges
- Source quality varies significantly — Synthesis Agent must distinguish authoritative sources from low-quality content using domain signals and recency.
- Agent pipelines can drift — each agent must receive a typed context contract to prevent downstream hallucination from context ambiguity.
- Cost management: Planner runs on a small model; only Synthesis and Writer use full GPT-4.
Results
- Produces citation-backed research reports on arbitrary topics in under 2 minutes.
- Demonstrates the orchestrator + specialist-agent pattern applied to information retrieval rather than task automation.