How astrodevil/adk-agent-examples works
Project Type and Purpose
• This is a collection of AI agent demos built on Google’s ADK (Agent Development Kit).
• It showcases five agents (Analyzer, Email, Sequential, Multi-Tool Search, Job Finder), each demonstrating ADK patterns like sequential pipelines, tool integration, and delegation.
File Structure
• Top level contains:
• 5 agent directories (analyzer_agent, email_adk_agent, seq_adk_agent, multi_tool_search_agent, jobfinder_agent)
• A shared requirements.txt, CLI entrypoints via adk commands, and environment setup files.
• Each agent folder contains:
• .env.example for API keys
• agent.py (or notebook in Job Finder) defining the ADK agent or pipeline
• README.md with agent-specific details
• No frontend code—interaction happens via ADK CLI, Dev UI (adk web), or FastAPI server (adk api_server).
README for Insights
• All agents use LiteLLM to integrate open-source LLMs (Meta-Llama, Nemotron Ultra, Qwen).
• External services (Nebius AI for inference, Exa/Tavily for search, Firecrawl for scraping, Resend for email, Mistral for OCR, Linkup for jobs) appear as tools.
• Patterns: sequential chaining, tool integration, agent delegation, multi-model approach, specialized agents.
a. Main Components
• User Interfaces:
• CLI (terminal)
• Dev UI (adk web)
• API Server (FastAPI via adk api_server)
• ADK Core/Orchestrator (central runtime)
• Agent Modules (5 separate pipelines):
• Analyzer Agent (5-step pipeline)
• Email Agent (single agent + Resend API tool)
• Sequential Agent (3-step pipeline)
• Multi-Tool Search Agent (root agent + delegated sub-agents)
• Job Finder Agent (4-step pipeline in notebook)
• LLM Integration Layer (LiteLLM)
• LLM Providers (Nebius AI endpoints for Meta-Llama, Nemotron Ultra; Qwen3-14B)
• External Tool Services: Exa, Tavily, Firecrawl, Resend, Mistral, Linkup
b. Relationships and Interactions
• Users → ADK CLI/UI/API → ADK Core
• ADK Core loads a selected agent module
• Within each agent:
• Sequential agents pass output from one sub-agent to the next
• Root agent delegates tasks to specialized sub-agents
• Agents invoke LiteLLM to call appropriate LLM provider
• Agents call external REST APIs (tools) to augment LLM results
• Final output returns via CLI, Dev UI, or API response
c. Architectural Patterns
• Sequential Pipelines: chain of agents (Analyzer, Sequential, Job Finder)
• Tool Integration: agents embed external API calls as “tools” (Email, Search, OCR)
• Agent Delegation: Multi-Tool Search agent delegates to sub-agents
• Multi-Model Approach: choose LLM per task
• Specialized Agents: each folder is a plug-and-play agent demo
d. Technologies/Frameworks
• Python, Google ADK (core runtime)
• LiteLLM for LLM connectivity
• FastAPI (via adk api_server)
• Various external RESTful services
Detail
• Agent’s internal flow if space allows (e.g., 5 sequential steps in Analyzer)
• Shared dependencies (requirements.txt, .env configs) as side notes
• Multi-model selection logic in LLM layer