Oracle Fusion AI Agent Studio Developer

Fusion agents work on business data. You learn how business objects and APIs expose it, and how to map objects to the data sources an agent can use, so that answers come from the right records.

• Fusion business objects and what they expose
• REST APIs for business data
• Mapping business objects to data sources
• Conversation topics and how they scope an agent
• Data access and what an agent should never see
• Testing a mapping with sample questions

You build A data map from three business objects to the sources and topics an agent will use.

AI Agent Studio offers more than one kind of agent. You learn how Workflow, Supervisor and Hybrid agents differ, choose among them for a process, and combine several into a team.

• What a Workflow agent does
• What a Supervisor agent does
• Hybrid agents and when to use them
• Choosing the agent type for a process
• Building a multi-agent team
• Hand-offs and responsibilities inside a team

You build An agent-type decision for a business process, with a team design where needed.

Tools are how an agent reads and acts. You define the tools it can call, write topics that route a conversation to the right agent, and write descriptions that help the model choose well.

• Kinds of agent tools
• Writing tool descriptions the model understands
• Conversation topics and routing
• Tool inputs, outputs and error cases
• Reusing tools across agents
• Keeping the tool set small and clear

You build An agent with a defined tool set and topics, tested on sample conversations.

An agent must serve only the people it is meant for. You apply role-based access control, work in the Security Console, and handle the extra permissions that advanced Workflow agent development needs.

• Role-based access control for agents
• Using the Security Console
• Separating who builds, runs and sees results
• Data security inside agent flows
• Permissions for advanced Workflow agents
• Testing access as different users

You build A role and permission design, tested by running the agent as different users.

Prompts decide how an agent behaves across many inputs. You write clear, user-centered prompts, build prompt flows that keep context from step to step, tailor them to business roles, and improve them from test results.

• Clear, user-centered prompt design
• Prompt flows across several steps
• Context retention within a flow
• Tailoring prompts to business roles
• Iterating from test results
• Keeping a prompt library

You build A prompt set for one agent, revised through two rounds of testing.

Workflow agents are built from nodes. You learn the node types, the patterns that recur in Fusion processes, and how to pass data between nodes without losing or leaking it.

• Workflow nodes and what each one does
• Branching, looping and conditions
• Passing data between nodes
• Common design patterns for business processes
• Error handling and fallbacks
• Keeping workflows readable

You build A Workflow agent that automates one business process end to end.

Agents that act need limits and links. You add guardrails and approval steps where a person must confirm an action, and connect external LLMs, MCP tools and REST services with care for security.

• Guardrails on inputs and outputs
• Human-in-the-loop approval steps
• Connecting REST tools
• Connecting MCP tools
• Using an external LLM
• Secrets, authentication and failure handling

You build An agent with guardrails, one approval step and one external connection.

An agent must be tested like any other software. You build test cases, run Agent Quality Evaluation, follow Agent Tracing through each step, and work through logs to find why a run failed.

• Designing test conversations and expected results
• Agent Quality Evaluation
• Agent Tracing: following a run step by step
• Reading logs to find the failing step
• Regression testing after a change
• Recording defects and fixes

You build A test set and a trace-based diagnosis of one failing agent.

A correct agent can still be slow or costly. You read prompt analytics and the debug window’s response time and token use, find the heavy steps, and trim them without losing quality.

• Prompt analytics: what to read
• Response time by step
• Token use and its causes
• Shortening prompts and context
• Simplifying tool sets and flows
• Re-testing after each optimization

You build An optimization report with before-and-after response time and token use.

Build one solution that uses everything above. You choose agent types, define tools and permissions, write the prompts, add guardrails and approvals, then test, trace and tune it.

• Choosing a Fusion process and writing its requirements
• Designing the agent team and its tools
• Building prompts, workflows and guardrails
• Setting roles and permissions
• Testing, tracing and tuning
• Presenting the solution and its limits

You build A multi-agent solution for one Fusion process, tested and traced, with a design record.