Deep Research Mode: How PROCUX Conducts Parallel AI Research in 90 Seconds

Introduction: The Research Time Problem

How long does comprehensive research take?

• Manual research (human): 2-4 hours

  • 30 min: Google searches, finding sources
  • 60 min: Reading 10-15 articles
  • 45 min: Taking notes, organizing information
  • 45 min: Writing summary/report

• A standalone AI chatbot (single-agent): 15-30 minutes

  • Limited to training data (can be outdated)
  • Limited or no live web search
  • Single perspective (no multi-agent analysis)

• PROCUX Deep Research Mode: 60-90 seconds

  • ✅ 50+ real-time web sources
  • ✅ 16 AI executives (parallel analysis)
  • ✅ Comprehensive report with citations
  • ✅ Real-time progress tracking

Why Traditional Research Fails

The Single-Agent Problem

A standalone AI chatbot (when used alone):

• ❌ Outdated data: bounded by a training cutoff
• ❌ No web search: can't reliably access real-time information
• ❌ Single perspective: only one AI viewpoint
• ❌ No sources: can't cite where information came from

Example Query: "What are the latest AI trends in enterprise adoption?"

Typical standalone-chatbot response:

"Based on my training data, AI trends include... [generic response based on older data]"

Problem: Trends change constantly, so a model bounded by an old training cutoff quickly goes stale.

The Manual Research Problem

Human researcher:

• ✅ Real-time data: Can Google and read latest articles
• ✅ Multiple sources: Can cross-reference 10+ sources
• ✅ Critical thinking: Can analyze and synthesize
• ❌ Slow: 2-4 hours for comprehensive research
• ❌ Limited scope: Can only read 10-15 articles
• ❌ Inconsistent: Quality varies by researcher

How PROCUX Deep Research Mode Works

PROCUX Deep Research Mode solves both problems with a 6-phase orchestrated pipeline that combines:

1. Multi-agent coordination (16 AI executives)
2. Real-time web search (live web search layer)
3. Content extraction (full-content extraction service)
4. Parallel execution (concurrent agent runs)
5. Knowledge synthesis (PROCUX Board™)
6. Report generation (Markdown/PDF)

Architecture Overview

User Query: "What are the latest AI trends in enterprise adoption?"
    ↓
┌─────────────────────────────────────────────────────────────┐
│ PHASE 1: PLANNING (0-15%)                                   │
│ - Select relevant agents (CEO, CTO, CMO)                   │
│ - Generate 5-10 sub-questions                              │
│ - Create search queries                                    │
│ Duration: 5-10 seconds                                     │
└─────────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────────┐
│ PHASE 2: WEB SEARCH (15-40%)                               │
│ - Execute 5-10 searches across the web                    │
│ - Retrieve 50+ sources (articles, reports, blogs)         │
│ - Rank by relevance                                       │
│ Duration: 10-15 seconds                                    │
└─────────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────────┐
│ PHASE 3: CONTENT SCRAPING (40-55%) - OPTIONAL             │
│ - Extract full text of top 5 sources                      │
│ - Extract full article content (HTML → Markdown)          │
│ - Enrich sources with detailed context                    │
│ Duration: 10-20 seconds                                    │
└─────────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────────┐
│ PHASE 4: AGENT ANALYSIS (55-85%)                           │
│ - Route sub-questions to relevant agents                  │
│ - 16 agents analyze in parallel (max 3 concurrent)        │
│ - Each agent generates findings with confidence scores    │
│ Duration: 25-35 seconds                                    │
└─────────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────────┐
│ PHASE 5: SYNTHESIS (85-95%)                                │
│ - Aggregate findings from all agents                      │
│ - Identify key themes and patterns                        │
│ - Calculate confidence scores                             │
│ - Generate executive summary                              │
│ Duration: 5-10 seconds                                     │
└─────────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────────┐
│ PHASE 6: REPORT GENERATION (95-100%)                       │
│ - Generate Markdown report with sections                  │
│ - Include source citations (APA/MLA format)               │
│ - Add agent insights and confidence scores                │
│ Duration: 5 seconds                                        │
└─────────────────────────────────────────────────────────────┘
    ↓
Final Report: Comprehensive 2,000-5,000 word research report
Total Time: 60-90 seconds

Phase 1: Research Planning

What happens: The system selects relevant AI executives and generates sub-questions.

Under the hood (simplified):

# PHASE 1: PLANNING (0-15%)
response.status = ResearchStatus.PLANNING
await update_progress(research_id, 5)

# 1. Select agents based on research type
agents = coordinator.select_agents(
    research_type=request.research_type,  # "market", "technical", "strategic"
    depth=request.depth,                   # "quick", "comprehensive", "exhaustive"
    include_agents=request.include_agents,
    exclude_agents=request.exclude_agents
)
response.agents_used = agents  # ["CEO", "CTO", "CMO"]

# 2. Generate research plan
plan = coordinator.generate_research_plan(
    topic=request.topic,
    research_type=request.research_type,
    depth=request.depth,
    agents=agents,
    language=request.language
)

# Plan includes:
# - sub_questions: ["What are emerging AI trends?", "Which industries adopt fastest?", ...]
# - search_queries: ["AI enterprise adoption 2025", "AI trends Q1 2025", ...]
# - target_agents: ["CEO", "CTO", "CMO"]
# - estimated_sources: 50
# - estimated_duration: 90 seconds

await update_progress(research_id, 15)

Example Output:

plan = ResearchPlan(
    sub_questions=[
        "What are the top AI trends in enterprise adoption for 2025?",
        "Which industries are adopting AI fastest and why?",
        "What are the main barriers to enterprise AI adoption?",
        "How do enterprise AI costs compare to traditional solutions?",
        "What ROI are companies seeing from AI investments?"
    ],
    search_queries=[
        "enterprise AI adoption trends 2025",
        "AI adoption by industry 2025 statistics",
        "barriers to enterprise AI implementation",
        "enterprise AI ROI case studies 2025",
        "AI vs traditional software costs comparison"
    ],
    target_agents=["CEO", "CTO", "CMO", "CFO"],
    estimated_sources=50,
    estimated_duration_seconds=85
)

Phase 2: Web Search (Real-Time Sources)

What happens: Parallel web searches retrieve 50+ real-time sources.

Under the hood (simplified):

# PHASE 2: WEB SEARCH (15-40%)
response.status = ResearchStatus.RESEARCHING
await update_progress(research_id, 20)

# Execute searches in parallel (batches of 3 to avoid rate limits)
search_results = await execute_searches(
    queries=plan.search_queries,
    max_results_per_query=10,
    search_depth="advanced"  # deeper, multi-pass search
)

# Convert to ResearchSource objects
all_sources = []
for result in search_results:
    sources = web_search.to_research_sources(result)
    all_sources.extend(sources)

response.sources_count = len(all_sources)  # ~50 sources
await update_progress(research_id, 40)

Web Search Layer Integration:

# The web search layer returns real-time sources in seconds
search_result = await web_search.query(
    query="enterprise AI adoption trends 2025",
    search_depth="advanced",  # More thorough than "basic"
    max_results=10,
    include_domains=["techcrunch.com", "venturebeat.com"],
    exclude_domains=["spam.com"]
)

# Returns (illustrative):
# - title: "Enterprise AI Adoption Keeps Climbing in 2025"
# - url: "https://example-tech-news.com/ai-adoption"
# - snippet: "A new industry report shows enterprise AI adoption continuing to rise..."
# - relevance_score: 0.92
# - published_date: "2025-02-15"

Example Sources Returned (50 total):

sources = [
    ResearchSource(
        title="Enterprise AI Adoption Keeps Climbing in 2025",
        url="https://example-tech-news.com/ai-adoption",
        snippet="A new industry report shows enterprise AI adoption continuing to rise...",
        relevance_score=0.92,
        published_date="2025-02-15"
    ),
    ResearchSource(
        title="Healthcare Among the Leaders in AI Adoption",
        url="https://example-healthtech.com/ai-healthcare",
        snippet="The healthcare sector is among the fastest adopters of AI...",
        relevance_score=0.88,
        published_date="2025-02-10"
    ),
    # ... 48 more sources
]

Phase 3: Content Scraping (Full-Text Extraction) - Optional

What happens: Scrape top 5 sources for full article content.

Under the hood (simplified):

# PHASE 3: CONTENT SCRAPING (40-55%) - OPTIONAL
if request.include_scraping and content_extractor.is_configured:
    await update_progress(research_id, 45)

    # Scrape top 5 sources
    top_urls = [s.url for s in all_sources[:5]]
    scraped = await content_extractor.scrape_batch(
        urls=top_urls,
        max_concurrent=3  # 3 pages at a time
    )

    # Enrich sources with full content
    for i, scrape_result in enumerate(scraped):
        if scrape_result.success and i < len(all_sources):
            all_sources[i].scraped_content = scrape_result.markdown
            all_sources[i].word_count = len(scrape_result.markdown.split())

    await update_progress(research_id, 55)

Why scraping matters:

• Without scraping: Only the 2-3 sentence snippet from search results
• With scraping: Full 1,000-3,000 word article content

Trade-off:

• Adds 10-20 seconds to research time
• Improves research depth significantly
• Optional (can disable for faster results)

Phase 4: Agent Analysis (Parallel Execution)

What happens: 16 AI executives analyze findings in parallel.

Under the hood (simplified):

# PHASE 4: AGENT ANALYSIS (55-85%)
await update_progress(research_id, 60)

# Prepare context for agents
context = {
    "topic": request.topic,
    "research_type": request.research_type,
    "sources_summary": [
        {"title": s.title, "url": s.url, "snippet": s.snippet[:200]}
        for s in sources[:10]  # Top 10 sources
    ],
    "user_context": request.user_context
}

# Create tasks for each agent
tasks = []
for agent_id in plan.target_agents:  # ["CEO", "CTO", "CMO", "CFO"]
    # Assign sub-questions to agents (round-robin)
    agent_questions = plan.sub_questions[i::len(plan.target_agents)]

    for question in agent_questions:
        tasks.append(
            coordinator.route_query_to_agent(
                query=question,
                agent_id=agent_id,
                context=context
            )
        )

# Execute in parallel (max 3 concurrent to avoid rate limits)
semaphore = asyncio.Semaphore(3)

async def limited_task(task):
    async with semaphore:
        return await task

results = await asyncio.gather(
    *[limited_task(task) for task in tasks],
    return_exceptions=True
)

# Collect findings
findings = []
for result in results:
    if isinstance(result, ResearchFinding):
        result.sources = sources[:3]  # Top 3 sources for this finding
        findings.append(result)

response.findings = findings
await update_progress(research_id, 85)

Example Agent Findings:

# Illustrative — the actual numbers come from the live sources retrieved
findings = [
    ResearchFinding(
        agent_id="CEO",
        question="What are the top AI trends in enterprise adoption?",
        finding="Three themes stand out in the retrieved sources: (1) generative AI "
                "adoption is accelerating, (2) multi-agent systems are moving into the "
                "mainstream, and (3) cost optimization is a top priority for buyers.",
        confidence=0.92,
        key_insights=[
            "GenAI adoption accelerating",
            "Multi-agent systems replacing single-model approaches",
            "Cost concerns driving demand for efficiency"
        ],
        sources=sources[:3],
        agent_name="Chief Executive Officer"
    ),
    ResearchFinding(
        agent_id="CFO",
        question="How do enterprise AI costs compare to traditional solutions?",
        finding="Sources indicate that, when well-scoped, AI initiatives can lower total "
                "cost of ownership versus traditional software — but poorly-scoped "
                "deployments can be more expensive, so optimization matters.",
        confidence=0.88,
        key_insights=[
            "Well-scoped AI can lower total cost of ownership",
            "Optimization is critical to realizing savings",
            "Payback depends heavily on the specific workload"
        ],
        sources=sources[:3],
        agent_name="Chief Financial Officer"
    ),
    # ... findings from CTO, CMO, etc.
]

Phase 5: Synthesis & Aggregation

What happens: Aggregate findings, identify themes, calculate confidence scores.

Under the hood (simplified):

# PHASE 5: SYNTHESIS (85-95%)
response.status = ResearchStatus.SYNTHESIZING
await update_progress(research_id, 90)

# Aggregate findings by agent, theme, confidence
aggregated = coordinator.aggregate_findings(
    findings=findings,
    research_type=request.research_type
)

# aggregated = {
#     "total_findings": 12,
#     "by_agent": {"CEO": 3, "CFO": 3, "CTO": 3, "CMO": 3},
#     "key_themes": [
#         "GenAI adoption acceleration",
#         "Multi-agent systems mainstream",
#         "Cost optimization priority"
#     ],
#     "confidence_avg": 0.89,
#     "all_sources": sources
# }

# Generate executive summary
response.summary = generate_summary(
    topic=request.topic,
    aggregated=aggregated,
    language=request.language
)

await update_progress(research_id, 95)

Phase 6: Report Generation

What happens: Generate comprehensive Markdown report with citations.

Example Report Structure (illustrative — the figures and sources shown are placeholders; a real run fills them in from the live sources retrieved):

# Research Report: Enterprise AI Adoption Trends 2025

**Research Date**: February 20, 2025
**Agents Consulted**: CEO, CFO, CTO, CMO
**Sources Analyzed**: 52

---

## Executive Summary

A concise, synthesized overview of what the retrieved sources say about the
topic — the leading trends, where sectors agree or diverge, and the main
drivers behind adoption — written in plain language for a decision-maker.

---

## Key Findings

### 1. GenAI Adoption Acceleration
**Source**: CEO Analysis

Each finding summarizes what the sources actually reported, with the drivers
the agent extracted, and links back to the specific pages it drew from.

**Supporting Sources**:
- [Source title](https://example-tech-news.com/ai-adoption)
- [Source title](https://example-industry-report.com/ai-investment)

---

### 2. Multi-Agent Systems Becoming Mainstream
**Source**: CTO Analysis

Multiple specialized AI agents working together are increasingly replacing
single-model approaches. Benefits the sources cite include cross-checking
between perspectives, more efficient context use, and role specialization
(a CEO agent for strategy, a CFO agent for finance, and so on).

**Supporting Sources**:
- [Source title](https://example-tech-news.com/multi-agent-ai)

---

### 3. Cost Optimization is a Priority
**Source**: CFO Analysis

Cost control is a recurring theme. The sources point to a few levers that
help — sending only relevant context, specializing agents, and caching
repeated calls — while cautioning that savings depend on the workload.

**Supporting Sources**:
- [Source title](https://example-industry-report.com/ai-cost)

---

## Agent Insights

Short, attributed takes from each executive that consulted on the topic —
CEO, CFO, CTO, CMO — each framed from that role's point of view, so the
reader sees the same evidence through several different lenses.

---

## Sources

Every source used is listed with its title, link, publisher, and date, so the
report is auditable end to end.

---

**Research Completed**: February 20, 2025
**Total Duration**: ~90 seconds

Real-World Performance

Example Scenario: Market Research for a Product Launch

Consider a startup preparing a fundraising pitch that needs comprehensive market research on a tight deadline. Commissioning a research firm can cost thousands of dollars and take weeks. Here is how the same work maps onto Deep Research Mode:

• Research topics (run as separate tasks):

  1. "AI-powered CRM market size and growth projections 2025-2030"
  2. "Competitor analysis: leading CRM platforms, strengths and weaknesses"
  3. "Enterprise CRM buyer personas and decision criteria"
  4. "CRM pricing strategies and market positioning"
  5. "AI CRM adoption barriers and objections"

• What you get: each task returns a structured, cited report in about a minute or two, drawing on dozens of live sources and several executive perspectives — so a founder can assemble a well-sourced market section in an afternoon instead of over weeks.

The value is speed and breadth: many sources, multiple viewpoints, and citations you can check — produced fast enough to keep up with a real deadline.

Deep Research Mode vs Alternatives

API Usage Examples

Basic Research

from procux import ProcuxClient

client = ProcuxClient(api_key="your_api_key")

# Start deep research
research_id = await client.start_deep_research(
    topic="AI enterprise adoption trends 2025",
    research_type="market",      # market, technical, strategic, competitive
    depth="comprehensive",        # quick, comprehensive, exhaustive
    include_scraping=True,        # Scrape top sources (adds 10-20s)
    max_sources=50
)

print(research_id)  # "550e8400-e29b-41d4-a716-446655440000"

# Get results
research = await client.get_research(research_id)

print(research.status)          # "completed"
print(research.summary)         # Executive summary
print(len(research.findings))   # 12 agent findings
print(len(research.sources))    # 52 sources
print(research.report_markdown) # Full report

# Export as PDF
pdf_url = await client.export_research_pdf(research_id)

Real-Time Progress Tracking

# Watch research progress in real-time
async def on_progress(research):
    print(f"Progress: {research.progress}%")
    print(f"Status: {research.status}")
    print(f"Sources found: {research.sources_count}")

research_id = await client.start_deep_research(
    topic="Healthcare AI adoption barriers",
    on_progress=on_progress  # Callback for progress updates
)

# Output:
# Progress: 5%  | Status: planning     | Sources: 0
# Progress: 15% | Status: planning     | Sources: 0
# Progress: 20% | Status: researching  | Sources: 0
# Progress: 40% | Status: researching  | Sources: 52
# Progress: 55% | Status: researching  | Sources: 52
# Progress: 60% | Status: analyzing    | Sources: 52
# Progress: 85% | Status: analyzing    | Sources: 52
# Progress: 90% | Status: synthesizing | Sources: 52
# Progress: 100%| Status: completed    | Sources: 52

Custom Agent Selection

# Research with specific agents
research_id = await client.start_deep_research(
    topic="AI cost optimization strategies",
    include_agents=["CFO", "CTO", "COO"],  # Only these agents
    exclude_agents=["CMO", "CHRO"]         # Skip marketing/HR perspectives
)

# Perfect for technical/financial research where marketing insights aren't relevant

Frequently Asked Questions

research_id = await client.start_deep_research(
    topic="Tendencias de adopción de IA empresarial 2025",
    language="es",  # Spanish
    search_language="es"  # Search Spanish sources
)

# Start research
POST /api/v2/deep-research
{
  "topic": "AI trends 2025",
  "research_type": "market",
  "depth": "comprehensive"
}

# Response
{
  "research_id": "550e8400...",
  "status": "pending"
}

# Get results
GET /api/v2/deep-research/550e8400...

# Response
{
  "research_id": "550e8400...",
  "status": "completed",
  "progress": 100,
  "summary": "...",
  "findings": [...],
  "report_markdown": "..."
}

const ws = new WebSocket('wss://api.procux.com/ws/research/550e8400...');

ws.onmessage = (event) => {
  const data = JSON.parse(event.data);
  console.log(`Progress: ${data.progress}%`);
  console.log(`Status: ${data.status}`);
};

research_id = await client.start_deep_research(
    topic="Enterprise AI trends",
    credentials={
        "nytimes": "your_api_key",
        "wsj": "your_api_key"
    }
)

# Save template
template_id = await client.create_research_template(
    name="Weekly AI Trends Monitor",
    topic="Latest AI trends and news this week",
    research_type="market",
    depth="quick",
    schedule="weekly"  # Auto-run every week
)

# Re-run manually
research_id = await client.run_research_template(template_id)

# Auto-scheduled research runs every Monday at 9 AM
# Results emailed to team@company.com

Conclusion: Dramatically Faster Research

The bottom line:

• ✅ Minutes, not hours — a comprehensive run finishes in roughly 60-90 seconds
• ✅ 16 AI executives analyze from multiple perspectives
• ✅ 50-80 real-time sources (vs none for a standalone chatbot, and only a handful for a person)
• ✅ Confidence scores and citations on every finding, so you can check the work
• ✅ A few cents per task instead of paying a researcher's hourly rate

Deep Research Mode isn't just fast—it's comprehensive, sourced, and affordable.

Next Steps

1. Try Deep Research: Start free trial (5 research tasks included)
2. See live demo: Watch research in action
3. Read the docs: Deep Research API
4. Explore templates: Research template library

Questions? Contact our team for a personalized research demo.

Related Resources

• PROCUX Board™: Multi-Agent Consensus Mechanism
• PROCUX Verify™: Evidence-Based AI
• Progressive Disclosure: Cost Optimization
• PROCUX DNA™: Company Learning System

Share this article: Twitter | LinkedIn | Facebook