Approach
Research agents fail because they confuse 'gathering information' with 'answering the question'. I built a two-phase agent: a planner that decomposes the question into verifiable sub-questions, then a searcher/synthesizer that answers each sub-question with cited evidence. The structured output schema forces the agent to commit to claims.
Problem
Generic research-agent demos sound impressive but produce unstructured walls of text that are hard to trust or re-use downstream.
How I built it
- ▸Question → planner → sub-question list.
- ▸Each sub-question → web search + synthesis with source tracking.
- ▸Final structured report with claims + citations in a fixed JSON schema.
Outcome
- →Reliable structured research reports.
- →Integrates into downstream automation workflows.
Stack
PythonClaudeGeminiWeb searchJSON schemas