Oracle Agentic AI in Integration 3

Overview

Oracle Agentic AI in Integration 3

Automate complex workflows using AI agents that intelligently orchestrate your Oracle integrations as tools — adapting in real time to achieve a defined business goal.

Traditional Integrations

Fixed, predefined sequence of steps
Runs exactly the same way every time
Best when you need full control over order
Ideal for enterprise connectivity
Complex when too many conditional paths

AI Agents

Adaptable — decides steps dynamically via LLM
Achieves a goal even when the path isn't predefined
Best when possibilities are too many to hard-code
Uses integrations as tools
Reasons, acts, observes and iterates

What's in this guide

Chapter 1

Key Concepts

Introduction to AI agents and how they fit with Oracle Integration

Chapter 2 — Tutorial 1

Build Your First Agent

Step-by-step: build an expense approval agent from scratch

Chapter 2 — Tutorial 2

Add Human Approval

Extend your agent to route decisions through a manager workflow

Chapter 3

MCP Server

Expose integrations as tools for third-party agent frameworks via MCP

Chapter 4

Design & Run Agents

Dynamic prompting, persistent run context, knowledge bases

Audience: Users who want to automate workflows with AI agents and agentic AI tools in Oracle Integration 3.

Key Concepts

Understand the building blocks before you start building agents.

Core building blocks

AI Agent

A software program that uses a Large Language Model (LLM) to reason and act — without human intervention — to achieve a specific goal. The agent decides which tools to call and in what order, based on what it learns from each step.

Agentic AI Tool

An integration that has been registered as a tool for AI agents. The agent invokes tools to interact with external systems. An integration must meet these requirements to become a tool:

Part of a project Status: Active REST Adapter trigger POST verb JSON request payload JSON response payload

Thinking patterns

ReAct (Reasoning + Action)

The agent alternates between reasoning and acting. It updates its plan based on new information at each step.

Think

Act

Observe

Repeat

Plan and Execute

The agent asks the LLM to generate a complete plan first, then executes each step sequentially.

Plan

Step 1

Step 2

…

You can also customise existing patterns or create entirely new ones for your agents.

Other key concepts

Human in the Loop

Enables human intervention at critical decision points. Use it when:

Approval is needed before executing a tool
The agent needs confirmation before proceeding
The agent doesn't know what to do next
A tool call results in an error needing human guidance

Prompt Templates

Ready-to-use system prompts with {{ payload.variable }} placeholders. Resolved from the request payload at runtime before being sent to the LLM. One agent can have multiple templates.

// Example Get the weather for {{ payload.city }} in USA Get the weather for {{ payload.city }} in FRANCE

Knowledge Bases (RAG)

Add corporate documents to a Retrieval-Augmented Generation knowledge base inside Oracle Integration. Agents and integrations can query these documents — turning static files into live, queryable knowledge.

Projects

All AI agents, integrations, and tools are created inside projects. Each project also acts as an MCP server, exposing its integrations as discoverable tools for third-party agent frameworks.

Tutorial 1

Build and Run Your First AI Agent

Build a simple expense approval agent. Expenses ≤ $50 are automatically approved; expenses > $50 are rejected. One integration, one tool, one agent.

What you need

Access to Oracle Integration (free trial available)
LLM base URL — e.g. https://api.openai.com
LLM model name — e.g. gpt-4o-mini
LLM API key

Steps at a glance

Step 1

Plan your agent

Define purpose, required tools, input/output parameters

Step 2

Create the project

Projects → Add → Create → name: "Simple Expense Approval Project"

Step 3

Create integration

Build "Auto Approve Expense Report" with REST trigger + POST + JSON

Step 4

Actions → Create agentic AI tool — write a clear description

Step 5

Configure thinking pattern

Create ReAct pattern; connect LLM (URL, model, API key)

Step 6

Build the agent

Set role + guidelines, attach tools, activate

Step 7

Run and test

Test Agent page → enter a natural language expense prompt

Step 8

Monitor runs

Observe tab → check all integration instances; review errors

Tool design

Tool name	Purpose	Input JSON	Output JSON
Auto Approve Expense Report	Automatically approves submitted expenses	`{ "AmountClaimed": "" }`	`{ "AutoApproveResponse": "" }`

Integration requirements for tool use

Part of a project Status: Active REST Adapter trigger POST verb JSON request payload JSON response payload

Mapper expression

// In the map action, wire AmountClaimed → AutoApproveResponse concat("Expense Report claimed for", /nstrgmpr:execute/ns21:requestwrapper/ns21:AmountClaimed, " is automatically approved")

Agent role & guidelines

Role: "You are an expense approval agent. Your role is to approve or reject expense claims." Guidelines: 1. The user submits a request for approval of expense claims. 2. If the expense is equal to or less than $50 USD, automatically approve the expense using the Auto Approve tool. 3. If the expense is more than $50 USD, reject the expense.

ReAct pattern guidelines

You should follow the ReAct agent pattern (Reason + Act). Think — internal reasoning Action — invoke tools to get information Observe — process tool response Repeat — until finished Do not guess or infer tool arguments except when: - Given specific guidance by the tool - You have previous context for a reasonable assumption

System integrations: When you configure a ReAct pattern, Oracle Integration auto-generates ReAct Pattern Intg and Get Tool Data Intg in your project. These are required — do not edit or delete them.

Monitoring a run

After running the agent: go to Observe to see all integration instances. Look for Auto Approve Expense Report, ReAct Pattern Intg, and Get Tool Data Intg. If any show errors, click the number in the Errored column → Instances tab → View Details.

Tutorial 2

Add Human Approval to Your Agent

Extends Tutorial 1: instead of rejecting expenses over $50, the agent routes them to a manager for human approval via a web form and workflow.

Two tools in this agent

Auto Approve Expense Report Request Human Approval

Prerequisites

Process Automation must be enabled on your Oracle Integration instance
Assign role ServiceDeveloper or ServiceAdministrator to users who need Human in the Loop access
Completed Tutorial 1 (or have an existing AI agent)

Build steps

Step 1

Create approval form

Human in the loop → Forms → Add. Name: "Simple Approval Form". Fields: Claim Amount, Approver Comments.

Step 2

Create approval workflow

Build a workflow with a start event, a user task (linked to the form), and an end event. Activate it.

Step 3

Create trigger connection

Create a REST trigger connection (OAuth 2.0 or Basic Auth) for the human approval integration.

Step 4

Create & activate integration

Build "Request Human Approval" integration: REST trigger + POST + JSON payload. Activate it.

Step 5

Actions → Create agentic AI tool. Write a clear description of when the agent should use this tool.

Step 6

Add tool + update agent

Add the human approval tool to the existing agent and update the agent guidelines to reference both tools.

Human approval flow

Agent receives expense

Is expense > $50?

Calls Human Approval tool

Manager sees web form

Approves or rejects

Agent receives response

Agent concludes

Updated agent guidelines

1. The user submits a request for approval of expense claims. 2. If the expense is equal to or less than $50 USD, automatically approve using the Auto Approve Expense Report tool. 3. If the expense is more than $50 USD, use the Request Human Approval tool to route to a manager for review.

After adding a new tool, always update the agent guidelines so the LLM understands when to invoke each tool. Clear, specific guidelines reduce LLM errors and improve accuracy.

Form fields

Field	Description
Claim Amount	The dollar amount the employee is claiming
Approver Comments	Free-text field for the manager to add notes on their decision

MCP Server

Use Integrations as Tools in an MCP Server

Each Oracle Integration project can act as an MCP (Model Context Protocol) server, allowing any third-party agent framework that supports MCP to discover and invoke your integrations as tools.

Supported frameworks

AI Agent Studio for Fusion Applications Langflow Any MCP-compatible framework

Workflow to use an integration via MCP

Step 1

Create client application

Create and activate the OAuth client application for authentication

Step 2

In your project, register the integration as an agentic AI tool with a clear, LLM-readable description

Step 3

Enable MCP for project

In project settings, switch on MCP server mode to expose registered tools

Step 4

Get MCP server URL

Copy the MCP server URL from the project — this is what MCP clients will connect to

Step 5

Discover tools from client

Configure your MCP client with the URL; it will auto-discover all registered integrations as callable tools

Key facts

Project = MCP server

Each project becomes its own MCP server with a dedicated URL. You choose which integrations within the project to expose.

Any integration, any framework

Any integration can be used as a tool regardless of which agent framework invokes it — Oracle Integration or third-party.

Tool descriptions are sent directly to the LLM. Write them precisely — the LLM uses the description to decide when and how to invoke each tool. Vague descriptions cause incorrect tool calls.

FAQ highlights

Question	Answer
Can I use a tool in both an Oracle agent and a third-party framework?	Yes. Once registered as an agentic AI tool, it's available via the MCP server URL to any client.
How many integrations can a project expose?	You selectively register which integrations to expose — you don't have to expose all of them.
Is the MCP URL stable?	Yes. The URL persists and can be shared with multiple clients.

Advanced Features

Go beyond the basics — dynamic prompting, persistent run memory, knowledge bases, and fine-tuned thinking pattern configuration.

Dynamic prompting

Prompt templates

Instead of a static system prompt, associate one or more prompt templates with your agent. Each template has placeholders that resolve from the incoming request payload at runtime.

// Template 1 Get the weather for {{ payload.city }} in USA // Template 2 Get the weather for {{ payload.city }} in FRANCE // Template 3 Get the weather for {{ payload.city }} in BRAZIL

The correct template is selected based on the payload context and sent to the LLM as the system prompt.

Persistent run context

Run agent with knowledge of previous runs

By default each agent run is stateless. You can configure the agent to carry context from previous runs — enabling it to build on past interactions, avoid repeated mistakes, and support multi-turn workflows.

Pass previous run context in the prompt payload
Useful for iterative approvals and multi-step workflows
Allows the agent to remember earlier decisions in a conversation

Knowledge bases

RAG knowledge bases

Upload corporate documents to a Retrieval-Augmented Generation knowledge base inside Oracle Integration. Agents can query these documents as a tool, and integrations can also query them directly.

Query from an integration Use as an agent tool Human-accessible reference

Thinking pattern configuration

Setting	Description	Notes
Max iterations	Limits the number of reasoning + action cycles before the agent provides a final answer	Prevents infinite loops
Temperature	Controls LLM creativity in reasoning	Higher = more creative; lower = more deterministic
Guidelines	Sent as the system prompt — defines how the agent reasons step by step	Be specific and structured
LLM connection	Base URL, model name, and API key for the LLM	Test connection before activating

Tool parameter visibility

Control what the LLM sees

When registering a tool, you control which parameters are visible to the LLM. For integrations with many fields, only mark as Visible the parameters that are relevant to the tool's stated purpose. This reduces noise and improves the LLM's ability to use the tool correctly.

Visible — sent to LLM as tool metadata Hidden — not sent to LLM

System-generated artifacts: Oracle Integration automatically creates ReAct Pattern Intg, Get Tool Data Intg, and an LLM Connection when you save a thinking pattern. These are required for agent operation — do not edit or delete them.

Using Agentic AI in Oracle Integration 3