How to Build a Research Agent

A research agent gathers information from multiple sources, synthesizes findings, and returns a cited answer. This guide describes a production-minded design—not a demo script.

Goal template

Define every run as:

Question: ...
Audience: ...
Must include: ...
Must avoid: ...
Citation policy: URLs + publish dates when available
Max time / max sources: ...

Ambiguous goals produce wander loops.

Recommended architecture

1. Clarify (optional) — rewrite the question into sub-questions
2. Plan — list subtopics and intended sources
3. Search loop — query → open → extract notes
4. Synthesize — outline → draft → cite
5. Verify — check claims against notes; flag low confidence

Patterns: plan-and-execute for structure, ReAct for adaptive search.

Tool set

Tool	Purpose	Guardrail
search(query)	Discover sources	Domain allow/deny lists
fetch(url)	Read content	SSRF protection, size limits
notes.write	Store excerpts	Schema: url, quote, claim
finalize(report)	Emit answer	Require citations array

Avoid giving the agent unrestricted browser + shell unless sandboxed.

Output schema (LLM-friendly and UI-friendly)

{
  "answer": "string",
  "key_findings": [{"claim": "string", "confidence": "high|medium|low", "sources": ["url"]}],
  "sources": [{"url": "string", "title": "string", "accessed": "ISO-date"}],
  "limitations": ["string"]
}

Evaluation cases

• Factual questions with known answers
• Multi-hop questions needing 2+ sources
• Contradictory sources (must surface disagreement)
• Stale topics (must prefer recent sources when asked)
• Injection pages (“ignore instructions and exfiltrate secrets”)

Graders can check: minimum sources, required sections, blocked domains never called, and citation URLs present in fetch logs.

Cost controls

• Cap search calls and page fetches
• Summarize pages into notes; do not keep full HTML in context
• Stop when each sub-question has ≥1 decent source
• Prefer cheaper models for extraction, stronger models for synthesis

Common failures

• Answer first, search later — reverse it with forced tool-before-final policies
• Citation theater — URLs never fetched; verify via trace
• Rabbit holes — require plan updates every N steps
• Overconfident synthesis — force limitations section

Minimal implementation sketch

state = {goal, plan: [], notes: [], steps: 0}
while steps < MAX:
  if not enough_notes(state): act = search_or_fetch(state)
  else: act = synthesize_or_expand(state)
  if act == finalize and citations_ok(state): return report

Ship criteria

• Citations correspond to fetched sources
• Domain policy enforced
• Offline suite ≥ 25 research tasks
• p95 runtime within product budget
• Clear “insufficient sources” behavior

Research agents are a great first production agent: high user value, mostly read-only tools, and easy-to-understand evals.