DeepResearchAgent (AgentOrchestra) Skill

A comprehensive assistant for working with DeepResearchAgent/AgentOrchestra - a state-of-the-art hierarchical multi-agent framework that achieved 83.39% accuracy on the GAIA benchmark for general-purpose task solving.

When to Use This Skill

This skill should be triggered when users ask about:

Core Framework:

• DeepResearchAgent or AgentOrchestra architecture
• Hierarchical multi-agent systems and coordination
• TEA Protocol (Tool-Environment-Agent) implementation
• Complex task decomposition and delegation strategies

Agent-Specific Questions:

• Planning Agent configuration and task decomposition
• Deep Researcher for information synthesis and research
• Browser Use Agent for web automation and data collection
• Deep Analyzer for extracting insights from complex data
• MCP Manager Agent for dynamic tool creation and evolution
• General Tool Calling integration and function invocation

Implementation Scenarios:

• Building multi-agent systems for complex problem solving
• Automating research workflows with multiple specialized agents
• Creating systems that need intelligent tool discovery and creation
• Implementing browser-based automation with multi-agent coordination
• Working with GAIA benchmark evaluation or similar complex tasks

Performance & Benchmarking:

• GAIA benchmark results (83.39% test accuracy, 82.4% validation)
• SimpleQA or HLE benchmark comparisons
• General-purpose LLM-based agent system design

Key Concepts

Hierarchical Architecture

AgentOrchestra uses a 2-layer hierarchical structure inspired by how a conductor orchestrates a symphony:

• Top Layer: Planning Agent - Central coordinator that decomposes complex objectives and delegates sub-tasks to specialized agents
• Bottom Layer: 6 Specialized Agents - Each with distinct capabilities for specific task types

The Six Specialized Agents

1. Planning Agent

   • Central coordinator for task decomposition
   • Delegates sub-tasks to appropriate specialized agents
   • Maintains overall workflow coordination

2. Deep Researcher Agent

   • Conducts thorough research on specified topics
   • Retrieves and synthesizes high-quality information
   • Supports multi-source information gathering

3. Browser Use Agent

   • Automates browser operations using Playwright/Patchright
   • Supports web search, information extraction, and data collection
   • Handles complex web interactions and navigation

4. Deep Analyzer Agent

   • Performs in-depth analysis of input information
   • Extracts key insights and identifies requirements
   • Supports multi-model analysis (Gemini 2.5 Pro)

5. General Tool Calling Agent

   • Provides general-purpose interface for tools and APIs
   • Supports function calling and external service integration
   • Handles standard automation tasks

6. MCP Manager Agent ⭐ NEW

   • Enables intelligent tool evolution through automated creation
   • Dynamic retrieval and systematic reuse of MCP tools
   • Addresses tool development inefficiency and version inconsistency

MCP Manager Agent Features

The MCP Manager Agent represents a paradigm shift from static tool provisioning to adaptive tool ecosystem management:

Core Capabilities:

• Keyword pre-filtering - Efficiently match tasks with relevant tools
• Automated tool generation - Create MCP-compliant tools through intent analysis
• Tool registry - Comprehensive registry with persistence and versioning
• Lifecycle tracking - Monitor tool usage and performance

Workflow:

1. Intent Analysis - Parse user task intentions and extract functional requirements
2. Tool Synthesis - Generate executable MCP-compliant tool implementations
3. Validation - Multi-stage evaluation for correctness and compatibility
4. Registration - Register validated tools with metadata and usage examples

Quick Reference

Agent Selection Guidelines

# Use Planning Agent for:
- Complex multi-step tasks requiring coordination
- Task decomposition and delegation strategies
- Overall workflow orchestration

# Use Deep Researcher for:
- Literature reviews and information synthesis
- Multi-source research requiring quality assessment
- Knowledge base construction and fact verification

# Use Browser Use Agent for:
- Web scraping and data collection
- Automated form filling and navigation
- E-commerce automation and monitoring

# Use Deep Analyzer for:
- Document analysis and insight extraction
- Pattern recognition in large datasets
- Requirements analysis from specifications

# Use MCP Manager for:
- Creating new tools for specific tasks
- Tool discovery and reuse optimization
- Managing evolving tool ecosystems

Benchmark Performance Context

GAIA Benchmark Results:
  Test Accuracy: 83.39%
  Validation Accuracy: 82.4%

Comparison Benchmarks:
  - SimpleQA: Evaluation on simple question-answering
  - HLE: Human-level evaluation for complex reasoning
  - Real-world tasks: Web search and reasoning combinations

Framework Integration Patterns

# Hierarchical Task Flow:
Planning Agent → Task Decomposition → Specialized Agent Selection → Execution → Result Synthesis

# Tool Evolution Workflow:
User Request → MCP Manager Analysis → Tool Creation/Discovery → Validation → Registration → Reuse

# Multi-Agent Coordination:
Complex Query → Planning Agent → Multiple Specialized Agents → Result Aggregation → Final Response

Reference Files

This skill includes comprehensive documentation in references/:

• architecture.md - Complete framework architecture, agent descriptions, MCP Manager details, benchmark results, and key principles for hierarchical multi-agent coordination

Use view references/architecture.md to access detailed technical documentation about the framework components and implementation strategies.

Working with This Skill

For Beginners

• Start with: Understanding the hierarchical architecture and six agent types
• Key concepts: Planning Agent as coordinator, specialized agents for specific tasks
• First steps: Learn when to use each agent type for different problem categories

For Framework Developers

• Focus on: MCP Manager Agent for dynamic tool creation and evolution
• Architecture patterns: Hierarchical coordination and TEA Protocol implementation
• Integration: Multi-agent workflow design and specialized agent coordination

For Researchers & Evaluators

• Benchmark context: GAIA results (83.39% accuracy) and comparison methodologies
• Evaluation approaches: SimpleQA, HLE, and real-world task assessment
• Performance analysis: Multi-agent system effectiveness measurement

For Implementation Teams

• Agent selection: Choose appropriate specialized agents for specific use cases
• Tool management: Leverage MCP Manager for automated tool creation and discovery
• Workflow design: Plan hierarchical task decomposition strategies

Advanced Features

MCP Tool Evolution

• Automated Creation: Generate MCP-compliant tools through intent analysis
• Dynamic Retrieval: Intelligent matching of tasks with existing tools
• Systematic Reuse: Comprehensive tool registry with versioning and lifecycle tracking

Multi-Model Support

• OpenAI: GPT-4.1, O3 series integration
• Anthropic: Claude 3.7 Sonnet support
• Google: Gemini 2.5 Pro for analysis tasks
• Local Models: vLLM integration with Qwen 2.5 series

Integration Capabilities

• Web Search: Google, Bing, DuckDuckGo, Baidu, Firecrawl
• Document Processing: PDF, DOCX, PPTX, Excel support
• Media Generation: Imagen (images) and Veo3 (video) integration
• Browser Automation: Playwright/Patchright for complex web interactions

Resources

Research & Documentation

• ArXiv Paper: "AgentOrchestra: A Hierarchical Multi-Agent Framework for General-Purpose Task Solving"
• GitHub Repository: Complete implementation and examples
• Website: https://skyworkai.github.io/DeepResearchAgent/

Academic Context

• Institution: Nanyang Technological University
• Domain: General-purpose task solving with LLM-based multi-agent systems
• Benchmark Suite: GAIA, SimpleQA, HLE evaluations

Notes

• AgentOrchestra represents a shift from simple dialogue to sophisticated multi-step reasoning
• The framework addresses LLM limitations in real-world environment interaction
• MCP Manager Agent enables adaptive tool ecosystem management vs. static tool provisioning
• Hierarchical design inspired by musical conductor orchestration principles
• State-of-the-art performance on general-purpose task solving benchmarks

Updating

To refresh this skill with updated documentation:

1. Re-run the scraper with the same configuration
2. The skill will be rebuilt with the latest information from the project website
3. Monitor GitHub repository for implementation updates and new examples