llm-agent-scaffold

llm-agent-scaffold

Generate a single-agent ADK 2.0 starter file using google.adk.agents.Agent (or LlmAgent).

When to use

• User says "create an ADK agent" / "scaffold a Gemini agent"
• User wants a quickstart agent.py they can run with adk web or adk run
• Single agent, no subagents, no workflow orchestration

When NOT to use

• Multi-agent setups → coordinator-subagent-pattern
• Sequential/Loop/Parallel patterns → workflow-agent-scaffold
• No-code YAML config → agent-config-no-code

Inputs to gather

• Agent purpose (e.g., "search assistant", "code reviewer")
• Model (default: gemini-2.5-flash; alternatives: gemini-2.5-pro, Claude via LiteLLM)
• Tools to bind (built-in like google_search, custom function tools, MCP tools)
• Instruction (system prompt) and one-line description

Output template

# agent.py
from google.adk.agents import Agent
from google.adk.tools import google_search

root_agent = Agent(
    name="search_assistant",
    model="gemini-2.5-flash",
    description="An assistant that can search the web for current information.",
    instruction=(
        "You are a helpful research assistant. Use the google_search tool "
        "to find authoritative information. Cite sources with URLs."
    ),
    tools=[google_search],
)

How to run

pip install google-adk --pre
adk web      # launches debug UI on localhost
adk run agent.py

Variants

• Switch Agent → LlmAgent for explicit base class (identical behavior in 2.0)
• Add output_key="result" to persist final output to session state
• Replace tools=[google_search] with custom function tools (see function-tool-author skill in adk-tools)

Validation

After generating, verify:

• Variable name is root_agent (ADK CLI convention)
• model string matches a current Gemini/Gemma/Claude identifier
• instruction is unambiguous and ≤ ~500 tokens