agent-config-no-code

agent-config-no-code

Author an ADK 2.0 agent as a YAML/JSON config file using Agent Config. Skip Python entirely.

When to use

• Domain expert wants to tweak prompts without touching code
• Agent definitions live in a config repo or CMS
• Rapid iteration on instruction/model without redeploying
• A/B testing agents declaratively

Template

# agent.yaml
name: customer_support
model: gemini-2.5-flash
description: Triages customer support tickets and routes to specialists.
instruction: |
  You are a tier-1 customer support agent. Classify each ticket as
  billing, technical, or general. For billing route to the billing_agent.
  For technical route to the technical_agent. Otherwise reply directly.
tools:
  - google_search
sub_agents:
  - $ref: ./billing_agent.yaml
  - $ref: ./technical_agent.yaml

Loading

from google.adk.agents import Agent

root_agent = Agent.from_config("agent.yaml")

Or via CLI:

adk run agent.yaml
adk web --config agent.yaml

Tools-only block

tools:
  - name: google_search
  - name: my_function_tool
    module: my_tools.weather
    function: get_weather
  - name: openapi_tool
    spec: ./petstore.openapi.yaml

Sub-agent reference

sub_agents:
  - name: billing_agent
    inline:
      model: gemini-2.5-flash
      instruction: Handle billing questions only.
  - $ref: ./shared/escalation_agent.yaml

Validation

• Top-level keys: name, model, instruction are required
• Tool refs must resolve at load time (importable modules + functions)
• $ref paths are relative to the config file
• Run adk validate agent.yaml (if available in your ADK version) before deploy