If you've been following this series, you already know Oracle AI Agent Studio can retrieve enterprise knowledge (shout-out to the Document Tool post), call external systems, and reason its way through open-ended questions. But there's a question I get asked a lot: "Can I trust an AI Agent to run an actual business process, not just chat?"
Turns out, Oracle already thought about this — by giving you two very different flavors of agent.
Two Agents, Two Personalities
Oracle AI Agent Studio ships with two agent types, and picking the right one is less "which is better" and more "which job are you hiring it for."
Workflow Agent (Deterministic) follows a predefined sequence of steps to automate structured business processes. AI only shows up at specific decision points, so the rest of the process stays predictable, repeatable, compliant, and easy to audit. Think approvals, document processing, order creation — anything where "consistent" beats "creative."
Supervisor Agent (Probabilistic) dynamically plans, routes, and coordinates work across multiple specialized agents based on the user's goal. Instead of following a fixed script, it figures out the best approach on the fly — a better fit for complex, cross-functional, ambiguous work like customer risk analysis or workforce planning.
Here's the analogy I keep coming back to: a Workflow Agent is a train on rails — it always arrives at the same stations, in the same order, and you know exactly where it'll stop. A Supervisor Agent is a taxi driver — same destination in mind, but it reroutes around traffic, picks its own streets, and improvises when Google Maps disagrees with reality. Neither is "smarter" than the other; they're built for different jobs.
| Workflow Agent | Supervisor Agent | |
|---|---|---|
| Behavior | Deterministic, fixed sequence | Probabilistic, dynamic routing |
| Where AI is used | Only at specific decision points | Throughout planning and execution |
| Best for | Approvals, document processing, order creation | Cross-functional, ambiguous, judgment-heavy work |
| Auditability | High — every step is traceable | Lower — path varies by run |
| Mental model | A train on rails | A taxi driver in traffic |
This post focuses on the Workflow Agent — because if you're automating something a compliance officer will eventually ask you to explain, "predictable" is a feature, not a limitation.
The Building Blocks
A Workflow Agent is assembled from five categories of nodes:
- AI Nodes – where the actual reasoning happens (agents, prompts)
- Logic Nodes – conditionals and branching
- Data Nodes – document processing and data transformation
- Workflow Control Nodes – orchestrating the sequence itself
- Communication Nodes – sending the output somewhere useful, like email
Rather than boil the ocean, let's build something small but real: an agent that answers benefits policy questions and emails the answer back. Basic on purpose — the goal is to get AI Nodes and Communication Nodes talking to each other correctly, before layering on logic and data nodes later in the series.
Step 1: Reuse the Document Tool, Add an Output Contract
The agent setup itself follows the same steps covered in the Document Tool post — attach a document tool so the agent grounds its answers in the actual benefits policy documents instead of guessing.
The twist this time: we define the output twice — once in the prompt, once in a formal output specification.
Prompt:
Return the result using the configured output specification:
- emailSubject: A concise subject line
- emailBody: The complete employee-facing response
- recipientEmail: The recipient's email address when available
Do not place the answer outside these output fields.
Output specification:
{
"type": "object",
"properties": {
"emailSubject": { "type": "string" },
"emailBody": { "type": "string" },
"recipientEmail": { "type": "string" }
},
"required": ["emailSubject", "emailBody"]
}
Why bother with a formal schema when the prompt already says the same thing in plain English? Because the prompt is a suggestion the model tries to follow, while the output specification is a contract the platform enforces — and, as you'll see in Step 2, it's the only thing standing between "this data flows automatically to the next node" and "this data doesn't exist as far as the workflow is concerned."
Step 2: Build the Workflow
With the agent ready, the workflow itself is refreshingly simple — just two nodes in sequence.
Step 2.1 — Halton Benefit Policy Agent Node
This is an AI Agent node, which can call domain specialist agents. Here, it's linked directly to the "Halton Benefits Policy Advisor WF" agent we configured in Step 1.
Step 2.2 — Halton Benefit Send Email Node
And here's exactly why the output specification from Step 1 matters. Oracle uses context-sensitive expressions to pass data between nodes, such as:
{{$context.$nodes.HALTON_BENEFIT_POLICY_AGENT_NODE.$input.message}}
{{$context.$nodes.HALTON_BENEFIT_POLICY_AGENT_NODE.$output.emailBody}}
These expressions only resolve to something useful because we explicitly defined emailSubject, emailBody, and recipientEmail back in Step 1. Skip the output specification, and that information never makes it to the email node — it's not that the workflow fails, it's that the fields simply won't be there to reference. Lesson learned: spec your outputs before you wire your nodes, or you'll be debugging an email that stubbornly refuses to say anything.
The next challenge is presentation — a plain-text email response reads like a support ticket auto-reply, not a "your company is on top of this" experience. The fix is to define the email body as an HTML template directly in the node, controlling font, color, and layout, while dropping in those same context expressions where the dynamic content belongs.
The Payoff
Running the agent and asking a benefits question produced exactly what we designed for: a formatted email containing the original question and the AI-grounded answer, styled and branded, landing straight in the inbox.
What's Next
This is deliberately a "basic" build — two nodes, one question, one email. A few honest limitations worth flagging before you take this to production:
- It sends one email per single answer right now, not a rolled-up summary across multiple questions.
- We've only touched AI Nodes and Communication Nodes; Logic, Data, and Workflow Control Nodes are still on the table.
That's exactly the point, though — get the wiring right on a small process before adding branching logic or multi-step document handling. More enhancements to this agent (and a deeper look at the other three node categories) are coming in future posts.
If you're building your own Workflow Agent, start with the same question I did: what's the one output my next node actually needs? Define that first, and the rest of the workflow gets a lot less mysterious.
Top comments (0)