Skip to main content

Deep Research Agent

Build an agent that researches topics deeply — searching the web, reading pages, and producing structured reports.

Overview​

A deep research agent combines:

  • Web search for discovering sources
  • Web browser for extracting page content
  • Summarization for distilling findings
  • RAG memory for accumulating knowledge across sessions

Quick Start​

# Scaffold with the research preset
wunderland init research-copilot --preset research-assistant
cd research-copilot

# Add the tools
wunderland extensions enable web-search
wunderland extensions enable web-browser

# Start chatting
wunderland chat

Then ask:

You: Research the current state of WebAssembly adoption in 2024-2025.
     Include major frameworks, browser support, and performance benchmarks.
     Produce a structured report with sources.

Library Setup​

import { createWunderland } from 'wunderland';

const app = await createWunderland({
  llm: { providerId: 'openai', model: 'gpt-4o' },
  preset: 'research-assistant',
  extensions: {
    tools: ['web-search', 'web-browser'],
  },
  skills: ['web-search', 'summarize', 'coding-agent'],
  discovery: {
    recallProfile: 'aggressive', // surface more tools when relevant
  },
  approvals: {
    mode: 'deny-side-effects', // read-only by default
  },
});

const session = app.session();
const result = await session.sendText(
  'Research quantum computing breakthroughs in the last 6 months. ' +
  'Focus on error correction advances. Cite all sources.'
);

console.log(result.text);

agent.config.json​

{
  "llmProvider": "openai",
  "llmModel": "gpt-4o",
  "personalityPreset": "analytical",
  "extensions": {
    "tools": ["web-search", "web-browser"]
  },
  "skills": ["web-search", "summarize"],
  "rag": {
    "enabled": true,
    "mode": "hybrid",
    "autoIngest": true
  },
  "security": {
    "preLlmClassifier": true,
    "dualLlmAuditor": true,
    "outputSigning": false,
    "riskThreshold": 0.7
  }
}

Why These Settings​

  • gpt-4o — Best reasoning for synthesizing multiple sources
  • analytical personality — Fact-focused, thorough, less creative embellishment
  • hybrid RAG — Stores research findings for follow-up sessions
  • autoIngest — Automatically extracts and stores key facts
  • deny-side-effects — Read-only tools (search, browse) work freely; writing requires approval

Research Workflow​

Single-Shot Research​

You: Research [topic]. Produce a structured report with:
     1. Executive summary
     2. Key findings (with source citations)
     3. Data points and statistics
     4. Open questions / areas for further research

Multi-Turn Deep Dive​

You: Search for recent advances in solid-state batteries.
Agent: [searches, returns initial findings]

You: Read the top 3 most relevant articles in detail.
Agent: [browses pages, extracts content]

You: Now synthesize everything into a technical brief.
Agent: [produces structured report with citations]

You: What are the main disagreements between researchers?
Agent: [analyzes sources for conflicting claims]

Comparative Analysis​

You: Compare React, Vue, and Svelte for a new enterprise dashboard project.
     Consider: performance, ecosystem, hiring pool, and long-term maintenance.
     Search for recent benchmarks and industry surveys.

Extension Stack​

ExtensionPurpose
web-searchMulti-provider search (Serper, SerpAPI, Brave)
web-browserPage content extraction, screenshot, structured data
news-searchNewsAPI integration for current events
image-searchFind relevant images and diagrams

Environment Variables​

# Required: at least one search provider
SERPER_API_KEY=...        # serper.dev (recommended)
# or
SERPAPI_API_KEY=...       # serpapi.com
# or
BRAVE_API_KEY=...         # search.brave.com

# Optional: for current news
NEWSAPI_API_KEY=...       # newsapi.org

Scaling Research​

RAG Memory for Ongoing Research​

With RAG enabled, your agent remembers findings across sessions:

# Session 1
You: Research the EU AI Act and its implications for startups.

# Session 2 (days later)
You: What did we learn about the EU AI Act last time?
Agent: [recalls from RAG memory] Based on our previous research...

You: Has anything changed since then? Search for updates.
Agent: [searches for new information, compares with stored knowledge]

Scheduled Research​

Combine with scheduling for automated research:

{
  "name": "weekly-industry-scan",
  "steps": [
    {
      "id": "search",
      "action": "web-search",
      "params": { "query": "{{topic}} news this week" }
    },
    {
      "id": "analyze",
      "action": "chat",
      "params": {
        "prompt": "Analyze these findings. Highlight anything that differs from what we knew before."
      }
    },
    {
      "id": "report",
      "action": "channel-post",
      "params": { "channel": "slack", "target": "#research" }
    }
  ]
}

Guardrails​

For research agents, keep these defaults:

  • Read-only by default — Search and browse don't modify anything
  • Require approval for posting — If the agent publishes findings, require human review
  • Verify sources — The analytical personality preset encourages citation
  • Rate limit searches — Avoid hitting API quotas in deep research sessions

Next Steps​