Designing a Hierarchical Deep Agent Architecture for Market Research

Introduction

Most AI applications today are little more than a single prompt sent to an LLM. While this works for simple tasks, it breaks down when the problem requires extensive research, validation, reasoning, and synthesis.

To solve this challenge, I built a multi-agent market research system capable of conducting deep industry analysis, gathering information from multiple sources, validating findings, and generating executive-level reports.

The architecture combines:

• GPT-4o as the primary reasoning engine
• DeepAgents for agent orchestration
• LangGraph for memory and execution state
• Tavily for web and news retrieval
• Specialized sub-agents for parallel research
• Retry mechanisms for production reliability

The result is a research workflow that resembles how a consulting team operates rather than how a traditional chatbot functions.

System Architecture

The system follows a hierarchical agent architecture.

User Query
                         │
                         ▼
                Lead Research Agent
                         │
       ┌─────────────────┼─────────────────┐
       │                 │                 │
       ▼                 ▼                 ▼

Financial       Competitive       Trend Research
Research         Analysis           Analysis

       ▼                 ▼                 ▼
                    Fact Checker
                         │
                         ▼
                    Synthesizer
                         │
                         ▼
                    Final Report

Instead of asking one model to perform every task, the workload is divided among domain-specific agents.

Each agent operates independently, performs focused research, and returns structured findings.

The final synthesis agent combines all outputs into a single report.

Persistent Knowledge Layer

One of the most important aspects of the architecture is the knowledge-loading mechanism.

agents_md
skills_md
instructions_md
examples_md

These files are loaded into memory and injected into the agent’s context.

This creates a lightweight knowledge base containing:

• Research methodologies
• Strategic frameworks
• Writing guidelines
• Output templates
• Internal operating procedures

Rather than hardcoding instructions inside prompts, the agent retrieves them from a centralized repository.

Benefits:

• Easier maintenance
• Better scalability
• Version-controlled prompt engineering
• Reusable across multiple projects

This pattern resembles Retrieval-Augmented Generation (RAG), but instead of retrieving business documents, the system retrieves operational knowledge.

Memory Management with LangGraph

The system uses LangGraph’s memory components:

InMemoryStore()
MemorySaver()

These solve two different problems.

InMemoryStore

Acts as a long-term knowledge repository.

Stores:

• Frameworks
• Templates
• Instructions
• Examples

This allows agents to access shared knowledge throughout execution.

MemorySaver

Acts as conversational state management.

Stores:

• Intermediate reasoning
• Agent outputs
• Previous interactions
• Workflow state

This becomes critical when reports require dozens of reasoning steps.

Without checkpointing, long-running workflows can lose context or exceed execution limits.

Tool-Augmented Research

The system doesn’t rely solely on model knowledge.

Instead, it uses external retrieval tools.

Web Search Tool

web_search()

Provides:

• Industry reports
• Company information
• Market data
• Analyst commentary

Parameters such as:

topic
time_range
search_depth
max_results

allow retrieval quality to be controlled.

News Search Tool

news_search()

Dedicated to recent developments.

Useful for:

• Product launches
• Acquisitions
• Funding rounds
• Earnings announcements

Separating general search from news search prevents the model from mixing evergreen information with rapidly changing events.

Dynamic Prompt Engineering

The system generates prompts dynamically using:

TODAY = date.today()
THIS_YEAR = date.today().year

These values are injected directly into prompts.

Example:

Include 2026 in every search query

Why?

Market research becomes stale very quickly.

By enforcing year-specific searches:

• Old statistics are avoided
• Current market conditions are prioritized
• Recency bias is controlled systematically

This is a simple but highly effective prompt-engineering strategy.

Multi-Agent Specialization

The most interesting aspect of the architecture is agent specialization.

Instead of one general-purpose researcher, multiple experts are created.

Example categories:

Financial Analyst

Responsible for:

• Revenue analysis
• Valuation metrics
• Earnings data
• Investor sentiment

Competitive Analyst

Responsible for:

• Market share
• Product comparisons
• Vendor positioning

Trend Analyst

Responsible for:

• Emerging technologies
• Growth signals
• Adoption patterns

Fact Checker

Responsible for:

• Verifying claims
• Cross-validating statistics
• Identifying inconsistencies

Synthesizer

Responsible for:

• Report generation
• Executive summaries
• Final recommendations

This mimics real-world consulting teams where specialists contribute domain expertise before findings are consolidated.

Parallel Execution

A major limitation of traditional agents is sequential execution.

For example:

Search A
Wait

Search B
Wait

Search C
Wait

This leads to high latency.

DeepAgents allows sub-agents to operate independently.

Conceptually:

Financial Research
Competitive Research
Trend Research
Fact Checking

can all run simultaneously.

Benefits:

• Lower response times
• Increased research coverage
• Better utilization of LLM resources

This pattern becomes increasingly important when workflows contain dozens of searches.

Reliability Through Retries

Production systems inevitably encounter:

• API rate limits
• Temporary failures
• Network instability

The architecture handles this using:

@retry(...)

with exponential backoff.

15s
30s
60s
90s

This approach provides several advantages:

• Reduces transient failures
• Prevents workflow termination
• Improves reliability under heavy load

Without retry logic, large research pipelines become fragile.

Hierarchical Prompt Design

The architecture uses two levels of prompts.

Global System Prompt

Defines:

• Research standards
• Validation requirements
• Citation rules
• Output quality expectations

This acts as the organization’s operating manual.

Agent-Specific Prompts

Each sub-agent receives:

• Shared context
• Domain responsibilities
• Specialized objectives

Example:

You are a competitive landscape researcher...

This ensures each agent remains focused and avoids task overlap.

The result is significantly better output quality than assigning all responsibilities to a single model.

Fact Verification Layer

One of the biggest problems in AI-generated research is hallucination.

To address this, the architecture introduces a dedicated verification agent.

Responsibilities include:

• Cross-referencing statistics
• Triangulating data
• Detecting inconsistencies
• Flagging unsupported claims

This creates a quality-control stage before final synthesis.

In practice, this dramatically improves trustworthiness.

Report Synthesis

After research is completed, outputs are merged by a synthesizer agent.

Responsibilities:

• Remove duplication
• Resolve conflicting findings
• Maintain consistent structure
• Generate executive-level recommendations

This final stage transforms fragmented research into a coherent strategic narrative.

Without synthesis, users receive disconnected information rather than actionable insights.

Key Engineering Patterns Demonstrated

This project showcases several advanced GenAI engineering concepts:

Agent Orchestration

Coordinating multiple specialized agents toward a shared objective.

Retrieval-Augmented Reasoning

Combining external search with LLM reasoning.

Long-Term Context Management

Maintaining frameworks, templates, and memory across workflows.

Parallel Research Pipelines

Reducing latency while improving research breadth.

Reliability Engineering

Handling failures through retries and checkpointing.

Fact Validation

Introducing verification layers to reduce hallucinations.

Dynamic Prompt Construction

Generating context-aware prompts based on time and task requirements.

Conclusion

The most interesting aspect of this system is that it moves beyond the traditional chatbot paradigm.

Instead of relying on a single model to answer a question, the architecture behaves more like a consulting organization:

• Researchers gather evidence
• Specialists analyze findings
• Fact checkers validate information
• Synthesizers produce recommendations

By combining DeepAgents, LangGraph, Tavily, and GPT-4o, the system demonstrates how modern AI applications can evolve from simple prompt-response interactions into scalable, multi-agent reasoning systems capable of producing high-quality research at enterprise scale.

This architecture can be extended far beyond market research into competitive intelligence, investment analysis, due diligence, policy research, strategic planning, and any domain requiring deep investigation and synthesis.

import os
from datetime import date
from pathlib import Path
from typing import Literal, Optional, List

from dotenv import load_dotenv
from tavily import TavilyClient
from tenacity import retry, wait_exponential, stop_after_attempt, retry_if_exception_type
from openai import RateLimitError

from deepagents import create_deep_agent
from deepagents.backends.store import StoreBackend
from deepagents.backends.utils import create_file_data
from langgraph.store.memory import InMemoryStore
from langgraph.checkpoint.memory import MemorySaver

load_dotenv("/Users/sandidas/Documents/Agentic_AI/.env")
os.environ["OPENAI_API_KEY"] = os.getenv("OPENAI_API_KEY")
os.environ["TAVILY_API_KEY"] = os.getenv("TAVILY_API_KEY")

TODAY = date.today().isoformat()
THIS_YEAR = date.today().year
PROJECTS = Path("/Users/sandidas/Documents/Agentic_AI/Deep Agents/Projects")
SKILLS   = PROJECTS / "skills"
agents_md       = (PROJECTS / "agents.md").read_text(encoding="utf-8")
skills_md       = (SKILLS   / "skills.md").read_text(encoding="utf-8")
instructions_md = (SKILLS   / "instructions.md").read_text(encoding="utf-8")
examples_md     = (SKILLS   / "examples.md").read_text(encoding="utf-8")
print(" Loaded:",
      f"agents.md={len(agents_md)} chars,",
      f"skills.md={len(skills_md)} chars,",
      f"instructions.md={len(instructions_md)} chars,",
      f"examples.md={len(examples_md)} chars")

store = InMemoryStore()
store.put(("memories",), "agents.md",       create_file_data(agents_md))
store.put(("memories",), "skills.md",       create_file_data(skills_md))
store.put(("memories",), "instructions.md", create_file_data(instructions_md))
store.put(("memories",), "examples.md",     create_file_data(examples_md))

tavily = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))

def web_search(
    query: str,
    max_results: int = 6,
    topic: Literal["general", "news"] = "general",
    time_range: Literal["day", "week", "month", "year"] = "year",
    search_depth: Literal["basic", "advanced"] = "advanced",
    include_domains: Optional[List[str]] = None,
):
    """Search the web with Tavily. ALWAYS include the year in the query for current data."""
    return tavily.search(
        query=query,
        max_results=max_results,
        topic=topic,
        time_range=time_range,
        search_depth=search_depth,
        include_domains=include_domains or [],
        include_answer=True,
    )

def news_search(query: str, days: int = 30, max_results: int = 6):
    """Breaking-news search — last N days only."""
    return tavily.search(
        query=query, topic="news", days=days,
        max_results=max_results, search_depth="advanced",
    )

SHARED_CONTEXT = f"""
TODAY IS {TODAY}.  CURRENT YEAR IS {THIS_YEAR}.
Treat any source older than 18 months as STALE unless foundational.
Prefer sources from the last 6 months for market-size, share, and trend claims.
Include "{THIS_YEAR}" in EVERY web_search query.

================================================================================
KNOWLEDGE 1 — PROJECT GUIDE (agents.md)
================================================================================
{agents_md}

================================================================================
KNOWLEDGE 2 — ANALYTICAL FRAMEWORKS (skills.md)
================================================================================
{skills_md}

================================================================================
KNOWLEDGE 3 — WORKFLOW & QUALITY RULES (instructions.md)
================================================================================
{instructions_md}

================================================================================
KNOWLEDGE 4 — OUTPUT FORMAT EXAMPLES (examples.md)
================================================================================
{examples_md}
"""

LIGHT_CONTEXT = f"""
TODAY IS {TODAY}.  CURRENT YEAR IS {THIS_YEAR}.
Include "{THIS_YEAR}" in EVERY web_search query.
Reject sources older than 18 months unless foundational.

================================================================================
FRAMEWORKS YOU MUST APPLY (skills.md)
================================================================================
{skills_md}

================================================================================
WORKFLOW & QUALITY RULES (instructions.md)
================================================================================
{instructions_md}
"""

SYSTEM_PROMPT = f"""You are an elite market research analyst for Cisco, operating
at the level of a top-tier strategy consultant (McKinsey / BCG / Gartner / Forrester).

{SHARED_CONTEXT}

================================================================================
HARD RULES (override anything else)
================================================================================
1.  Always include "{THIS_YEAR}" or "{THIS_YEAR-1}" in EVERY web_search query.
2.  Run AT LEAST 8–12 searches before writing the final report.
3.  Use `news_search` for earnings, product launches, M&A in the last 90 days.
4.  Apply frameworks from skills.md (TAM/SAM/SOM, Porter's 5F, JTBD, SWOT).
5.  Match the format and citation style from examples.md exactly.
6.  Cite EVERY claim with [n] inline + a Sources block (URL + publish date).
7.  Triangulate every material number across 3+ independent sources.
8.  Reject sources older than 18 months unless foundational; label stale stats.
9.  Never invent statistics, companies, quotes, or executives.
10. End with a "Strategic Implications" section (3 specific recs for Cisco).
11. Target 1,800–2,500 words. Depth beats brevity.
12. Delegate research to sub-agents, then synthesize their findings.
"""

def build_subagent(name: str, role: str, focus: str) -> dict:
    return {
        "name": name,
        "description": f"{role} {focus}",
        "system_prompt": f"""You are a {role}

{LIGHT_CONTEXT}

================================================================================
YOUR SPECIFIC FOCUS
================================================================================
{focus}

RULES:
- Include "{THIS_YEAR}" in every search query.
- Run 3–5 searches before answering.
- Apply the relevant frameworks from skills.md.
- Cite every claim with [n] + a Sources block (URL + date).
- Return a ≤400-word findings summary.
""",
        "tools": [web_search, news_search],
        "model": "openai:gpt-4o-mini",
    }

subagents = [
    build_subagent(
        "StockMarketResearcher",
        "Stock-market and financial-data specialist.",
        f"Pull the latest 10-K/10-Q numbers, EV/Revenue multiples, and analyst price "
        f"targets for Cisco, Arista, Juniper, HPE/Aruba, Palo Alto, Fortinet, NVIDIA "
        f"networking. Use SEC filings and earnings transcripts dated {THIS_YEAR}.",
    ),
    build_subagent(
        "StrategyArchitect",
        "Strategy framework expert.",
        "Apply Porter's 5 Forces, TAM/SAM/SOM, and SWOT to the AI-networking market "
        "with Cisco as the anchor. Return numbers + framework outputs in tables.",
    ),
    build_subagent(
        "CompetitiveAnalyst",
        "Competitive-landscape researcher.",
        f"Map top 7 networking vendors: market share (IDC, Dell'Oro, Synergy {THIS_YEAR}), "
        "product portfolio, recent launches, pricing posture, partnerships.",
    ),
    build_subagent(
        "TrendsAnalyst",
        "Quantitative trend researcher.",
        "Quantify 5 trends: AI fabric / RoCE, AIOps, SASE/SSE, 800G optics, "
        "hyperscaler capex. Use search/funding/hiring signals as evidence.",
    ),
    build_subagent(
        "FactChecker",
        "Technical fact-checker.",
        "Re-verify EVERY numeric claim across 3+ independent sources. Flag any "
        "number that cannot be triangulated. Return a verification table.",
    ),
    build_subagent(
        "Synthesizer",
        "Senior synthesizer / deck designer.",
        "Merge sub-agent outputs into the final brief in the EXACT format shown "
        "in examples.md. Preserve all citations.",
    ),
]

agent = create_deep_agent(
    model="openai:gpt-4o",                    
    tools=[web_search, news_search],
    backend=StoreBackend(store=store),        
    system_prompt=SYSTEM_PROMPT,
    checkpointer=MemorySaver(),
    subagents=subagents,
)

USER_PROMPT = f"""TODAY IS {TODAY}. Produce a market-research brief for Cisco's
exec team on the AI-driven enterprise networking market. Cover:

1. Market size & growth ({THIS_YEAR} → 2028, with TAM/SAM/SOM and CAGR).
2. Top 7 competitors with the LATEST quarterly financials, product moves, posture.
3. Buyer personas + top 5 purchase criteria (G2 / Gartner Peer Insights {THIS_YEAR}).
4. 5 key trends quantified with {THIS_YEAR-1}–{THIS_YEAR} evidence.
5. Strategic implications + 3 specific 12-month recommendations for Cisco.

Delegate to your sub-agents in PARALLEL where possible. Have the Synthesizer
assemble the final brief in the EXACT format from examples.md. Target 2,000–2,500 words.
"""

@retry(
    retry=retry_if_exception_type(RateLimitError),
    wait=wait_exponential(multiplier=2, min=15, max=90),   # 15s → 30s → 60s → 90s
    stop=stop_after_attempt(5),
    reraise=True,
)
def safe_invoke(payload, cfg):
    return agent.invoke(payload, config=cfg)

result = safe_invoke(
    {"messages": [{"role": "user", "content": USER_PROMPT}]},
    {
        "configurable": {"thread_id": f"cisco_brief_{TODAY}"},
        "recursion_limit": 100,
    },
)

final = result["messages"][-1].content
if isinstance(final, list):
    for block in final:
        print(block.get("text", block) if isinstance(block, dict) else block)
else:
    print(final)