Building a Deep Research Agent
Research assistants handle multi-step tasks requiring information from multiple sources.
The Research Pipeline
- Query Decomposition: Break question into 3-5 sub-questions.
- Parallel Search: Search multiple sources simultaneously.
- Source Evaluation: Score sources for relevance and reliability.
- Synthesis: Combine findings with citations.
- Verification: Cross-reference all claims against sources.
Agent Architecture
const researchPipeline = {
decomposer: { model: "sonnet", task: "Break into sub-questions" },
searcher: { model: "haiku", tools: ["web_search", "arxiv"], parallel: true },
synthesizer:{ model: "sonnet", task: "Write analysis with citations" },
verifier: { model: "haiku", task: "Verify claims against sources" }
};
Key Design Patterns
| Pattern | Implementation | Benefit |
|---|
| Query Decomposition | Break complex Q into simple Qs | Better search results |
| Parallel Search | All sub-queries searched at once | 3-5x faster |
| Source Scoring | Rate by authority + recency | Filters noise |
| Citation Verification | Cross-reference claims | Eliminates hallucinated citations |
🎯 Pro Tip: Citation verification is NON-NEGOTIABLE. Without it, the agent WILL hallucinate citations. Use a cheap model to cross-reference every claim.
Cost Profile (typical research task)
| Step | Model | Cost |
|---|
| Decomposition | Sonnet | $0.02 |
| Search (5 sub-q × 5 sources) | Haiku | $0.01 |
| Synthesis | Sonnet | $0.06 |
| Verification | Haiku | $0.004 |
| Total | | ~$0.10 |