Research agent

A research agent takes a question, searches the web (and internal sources), reads results, synthesizes a structured answer with citations. Useful for competitive intelligence, market research, due diligence, fact-finding.

The loop

1. Decompose the question into sub-queries
2. Search for each sub-query
3. Read top results
4. Extract relevant claims with source links
5. Identify gaps; iterate with more searches if needed
6. Synthesize final answer with citations

Core tools

• web_search(query), public web search
• read_url(url), fetch and parse a page
• search_internal(query), internal knowledge base
• summarize(text, focus), distill relevant info

Design choices

Breadth vs depth

Breadth: more queries, shallower read per result. Depth: fewer queries, full-article read per source. Research tasks usually need both in sequence.

Source diversity

Don't summarize from one source. Require multiple corroborating sources for key claims.

Citation quality

Every factual claim should link to the specific source. Without citations, hallucinations creep in.

Failure modes

• Research keeps going, never produces final output (budget hits first)
• Summarizes from a single source without cross-checking
• Cites sources that don't actually support the claim
• Misses a key perspective (no counter-search)

Production tools

Frameworks: Manus, OpenDevin (code research), Claude with web search, custom pipelines built on LangGraph. Commercial: Perplexity, Consensus, Elicit.