oracle-adk

Oracle ADK Expert Skill

Purpose

Master Oracle's Agent Development Kit (ADK) for building enterprise-grade agentic applications on OCI Generative AI Agents Service with code-first approach and advanced orchestration patterns.

Platform Overview

OCI Agent Development Kit (Released May 22, 2025)

Client-side library that simplifies building agentic applications on top of OCI Generative AI Agents Service.

Key Value: Code-first approach for embedding agents in applications (web apps, Slackbots, enterprise systems).

Requirements: Python 3.10 or later

Core Capabilities

1. Multi-Turn Conversations

Build agents that maintain context across multiple interactions.

Pattern:

from oci_adk import Agent

agent = Agent(
    name="customer_support",
    model="cohere.command-r-plus",
    system_prompt="You are a helpful customer support agent"
)

# Multi-turn conversation
conversation = agent.create_conversation()
response1 = conversation.send("I need help with my order")
response2 = conversation.send("It's order #12345")
# Agent remembers context from previous messages

2. Multi-Agent Orchestration

Routing Pattern:

# Route requests to specialized agents
def orchestrator(user_query):
    if requires_technical_support(user_query):
        return technical_agent.handle(user_query)
    elif requires_billing(user_query):
        return billing_agent.handle(user_query)
    else:
        return general_agent.handle(user_query)

Agent-as-a-Tool Pattern:

# One agent uses another agent as a tool
main_agent = Agent(
    name="supervisor",
    tools=[research_agent, analysis_agent, report_agent]
)

# Main agent orchestrates specialist agents
result = main_agent.execute("Research and analyze Q4 performance")

3. Deterministic Workflows

Build predictable, orchestrated workflows with explicit control flow.

from oci_adk import Workflow, Step

workflow = Workflow([
    Step("validate_input", validation_agent),
    Step("process_request", processing_agent),
    Step("generate_response", response_agent)
])

result = workflow.execute(user_input)