Deep Research Agent

A multi-agent research system built with Kotlin and Koog Agents that conducts comprehensive research through intelligent agent coordination and web search.

🏗️ Architecture

Three-phase pipeline with specialized AI agents:

SCOPING → RESEARCH → REPORT GENERATION

Agents:

• ClarificationAgent: Interactive questioning to refine research scope
• BriefGenerationAgent: Converts user input into detailed research briefs
• ResearcherAgent: Individual research with web search capabilities
• SupervisorAgent: Coordinates multiple researchers and parallel research
• FinalReportAgent: Synthesizes findings into comprehensive reports

Key Features:

• Parallel research execution (up to 3 concurrent tasks)
• Tavily web search integration with source attribution
• Structured report generation with citations
• Built-in evaluation framework

🧠 Agent Hierarchy & Planning

Hierarchical Structure

The system uses a supervisor-subagent architecture for specialized research:

SupervisorAgent (Coordinator)
├── ResearcherAgent 1 (Context A)
├── ResearcherAgent 2 (Context B)  
└── ResearcherAgent 3 (Context C)

Supervisor Agent:

• Analyzes research brief and breaks it into specialized subtopics
• Decides between single vs. parallel research based on complexity
• Each subtopic gets independent context and instructions
• Aggregates findings from all sub-agents

Researcher Agents:

• Work independently with focused context
• Execute parallel research on different aspects
• Use ReAct pattern: Reason → Act → Observe → Plan
• Built-in reflection after each search to plan next steps

ReAct Planning Pattern

Each researcher follows a Reasoning-Acting cycle:

1. Reason: Analyze current findings and identify gaps
2. Act: Execute targeted web searches with specific queries
3. Observe: Process search results and extract key information
4. Plan: Reflect on progress and determine next actions

This creates deliberate, strategic research rather than random searching.

🚀 Installation

Prerequisites

• Java 21+
• OpenAI API Key
• Tavily API Key

Setup

1. Clone and build

git clone <repository-url>
cd deep-research-agent
./gradlew build

2. Set environment variables

export OPENAI_API_KEY="your-openai-api-key"
export TAVILY_API_KEY="your-tavily-api-key"

3. Run

./gradlew run

⚙️ Configuration

Research Parameters

val researchWorkflow = ResearchWorkflow(
    apiKey = openaiKey,
    tavilyApiKey = tavilyKey,
    maxConcurrentResearchers = 2,  // Parallel research limit
    maxIterations = 20             // Maximum research iterations
)

Search Settings

TavilySearchTool.Args(
    query = "research topic",
    searchDepth = "advanced",      // "basic" or "advanced"
    maxResults = 5,                // 1-10 results
    includeRawContent = true,      // Include full content
    topic = "general"              // "general" or "news"
)

🧪 Evaluation

Run the built-in evaluation:

val scopeEvaluator = ScopeEvaluationRunner(openaiKey)
val passed = scopeEvaluator.run()