paths: agent/agents/**/*.py

Google ADK Agent Implementation Rules

Agent Structure

• Agents extend LlmAgent or BaseAgent
• Define name as unique identifier
• Use description for orchestrator routing decisions
• Keep instruction focused and actionable

State Management

• Use output_key for state passing between agents
• Access state via ctx.session.state.get("key")
• Set state via ctx.session.state["key"] = value
• Use unique keys to avoid collisions in parallel execution

LlmAgent Pattern

agent = LlmAgent(
    name="unique_name",
    model="gemini-2.5-flash",  # or LiteLlm(model="provider/model")
    description="Clear description for routing",
    instruction="Specific instructions for the agent",
    tools=[...],  # MCP tools or FunctionTools
    output_key="result_key"  # For state passing
)

Composite Agents

SequentialAgent

• Order matters: first to last execution
• Each sub-agent can access previous outputs via state
• Minimum 2 sub-agents required

ParallelAgent

• All sub-agents run concurrently
• Each MUST have unique output_key
• Results collected after all complete

LoopAgent

• Set max_iterations to prevent infinite loops
• Include termination condition in sub-agent logic
• Use for iterative refinement patterns

Custom Agents

• Extend BaseAgent for custom logic
• Implement _run_async_impl(self, ctx) method
• Yield events from sub-agents when delegating
• Use for conditional routing or complex workflows

Tool Integration

• Register MCP tools via FunctionTool.from_mcp_tool()
• Use @FunctionTool decorator for Python functions
• Include clear parameter descriptions
• Handle tool errors gracefully

Multi-Model Support

# Native Gemini
model="gemini-2.5-flash"

# Via LiteLLM
from google.adk.models.lite_llm import LiteLlm
model=LiteLlm(model="openai/gpt-4o")
model=LiteLlm(model="anthropic/claude-3-haiku-20240307")
model=LiteLlm(model="ollama_chat/llama3.2")

Testing Agents

• Test agents in isolation before composition
• Mock MCP tools for unit tests
• Use InMemorySessionService for testing
• Verify state changes after execution