How to build an agent team in Oracle AI Agent Studio

An agent team is the unit you deploy from Oracle AI Agent Studio: one or more agents with topics, instructions and tools, tested together and published into a Fusion page or job. This guide walks through building one, from choosing between a seeded template and a blank team to scoping tools, adding human approval, testing and publishing, with a hypothetical quote-to-requisition example.

Reviewed 10 October 2026 8 min read

On this page

The parts of an agent team and what each controls
Copy a template, use a template or start from a blank team
Writing topics and instructions for an ERP task
Tool types and the scoping question for each
From blank canvas to published agent team
A hypothetical quote-to-requisition assistant
Pre-release checks for a new agent team
Questions and answers
Sources

The parts of an agent team and what each controls

AI Agent Studio uses a small vocabulary. Getting it straight first avoids configuring the right behavior in the wrong place¹.

Agent team: The deployable unit: a single agent or a group of agents following a structured sequence to complete a business task or answer a question.
Agent: Uses a language model to reason, plan actions and talk to users. Common roles are a supervisor that orchestrates other agents and specialists skilled with particular tools.
Topic: An area of expertise with boundaries, defined through instructions that constrain what the agent discusses and does.
Instructions: Natural-language rules attached to a topic and sent to the model as part of its prompt, including guidelines and guardrails.
Tool: A capability an agent can call, such as a business object tool for Fusion records or a document tool for retrieval. Tools are reusable across agents.
Testing: The built-in workspace where administrators check tool, topic and instruction configuration and inspect the agent's reasoning and cited sources.

Copy a template, use a template or start from a blank team

Oracle offers two ways to start from a preconfigured template, and you can also build your own team. Preconfigured agents, tools and topics cannot be edited directly; you change them by copying the artifact and adding the copy to your team².

If
A seeded template covers the process and you mainly want to adjust wording, tools or scope.
Then
Use Copy Template: it adds your suffix to the team and its editable artifacts, then opens the canvas for editing³.
You inherit Oracle's tested design and keep a naming trail back to the original.
If
You want to rename and configure the team's general settings before you see the canvas.
Then
Use Use Template, which starts with general settings and opens the canvas once the team exists³.
The difference between the two options is the order of renaming and editing, not what you can change.
If
No template fits, or the template's scope is much broader than the task.
Then
Build a custom team with one narrowly scoped agent, adding agents only when a second specialist is clearly needed.
A small team is easier to test, secure and explain to the people who approve its output.

Writing topics and instructions for an ERP task

Topics and instructions decide most of an agent's behavior, so write them the way you would brief a new colleague on a controlled process. Name the task precisely ('create draft purchase requisitions from supplier quotations'), not the department ('help procurement'). State the inputs the agent can rely on and the fields it must never guess, such as cost center, supplier site or tax code.

Then write the refusals and handovers. An ERP agent should say what it cannot do, stop when a required value is missing and pass the case to a person with what it found. Ask it to cite the records it used and to describe its output accurately: 'I created a draft requisition for you to review', never 'your order is placed'. Keep each instruction short and testable, so it maps to a test case you can run later.

Do not restate business rules that Fusion already enforces. Thresholds, tolerances and validations belong in the application, where they apply whatever the model does; instructions explain how to work within them.

Tool types and the scoping question for each

The tool list grows with each release, so confirm what your environment offers. Oracle's documentation describes the types below¹, and its October 2025 announcement added Model Context Protocol support for external tools⁴.

Tool type	What it gives the agent	Scoping question before you attach it
Business object	Retrieve, create, update or delete Fusion records through business objects	Which objects and which of those actions does this task need, and does the tool act with the user's own data access?
Document retrieval	Answers grounded in documents you load, such as policies or contract terms	Which documents are authoritative, who keeps them current and who is allowed to see them?
Email	Sending messages as a step in a flow	Should sending require approval, and which recipients are allowed?
Connector or external REST	Calls to systems outside Fusion	Which credential is used, can it write data, and would a write bypass Fusion approvals?
MCP	Tools exposed by an external Model Context Protocol server	Who operates the server, which tools does it list and where are its credentials stored?
Calculator and user query	Arithmetic, and structured questions back to the user	Would a Fusion rule that auditors already rely on do this calculation better?

Tool names and availability vary by Fusion release, so check the documentation for your update.

From blank canvas to published agent team

List what Oracle already provides
Check the seeded agents and templates in your release for this process; extending one usually leaves fewer artifacts to maintain than building a parallel team.
Output
Template shortlist
Owner
Fusion administrator
Create the team and name it for its job
Copy or use the template, apply a suffix your organization will recognize, and record which seeded team it came from and why.
Output
Draft agent team
Owner
Fusion administrator
Define agents and their roles
Keep one agent unless the task genuinely splits. If it does, let a supervisor route the work and give each specialist only the tools its part needs.
Output
Agent roster with roles
Owner
Implementation lead
Attach topics and write instructions
Write boundaries, required inputs, refusals and handovers as described above, and link each instruction to a test case.
Output
Topic and instruction set
Owner
Process owner with implementation lead
Attach tools with least privilege
Add only the tool actions the task needs. Prefer read access until a write is approved, and record the credential behind every external tool.
Output
Tool access register
Owner
Security administrator
Add the approval step
In a workflow, place a human approval node after the agent has prepared its output and before any record is created or sent. From release 26C a node can use an approval process with levels, approver rules and conditional routing⁵.
Output
Approval design
Owner
Process owner
Choose the model
Where your release allows it, select the language model the team uses, and run the same test cases on any alternative before switching. Oracle supports models from several providers in AI Agent Studio⁴.
Output
Model decision record
Owner
Implementation lead
Test against real exceptions
Run the clean cases and the awkward ones: missing fields, duplicates, ambiguous suppliers, out-of-policy requests. Inspect reasoning and cited sources in the testing workspace, and fix instructions rather than individual outputs.
Output
Test log with outcomes
Owner
Process owner and testers
Secure, publish and monitor
Restrict the team to the roles that should use it, publish the version you tested and agree who reviews usage and errors after go-live.
Output
Published team with a support owner
Owner
Fusion administrator

A hypothetical quote-to-requisition assistant

Hypothetical example

Extending a seeded agent instead of building one

Oracle's Quote to Purchase Requisition Assistant reads a supplier quotation PDF forwarded to a configured mailbox, maps it to requisition fields and creates a draft requisition for the requester to check and submit. It handles text-based PDFs, not scanned images⁶.

For example, a mid-sized manufacturer might copy that template with its own suffix, add a topic that refuses quotes from suppliers without an active site in Fusion, and add an instruction to leave the charge account empty when a quote names no project, so the requester supplies it.

Its testers would run recent real quotations, including multi-currency and multi-page ones, check that a quote forwarded twice does not create a second draft, and confirm that requisition approval rules still apply after submission.

Pre-release checks for a new agent team

0 of 7 checked

Every tool action is justified in writingEach create, update or delete permission maps to a named step in the process.
Refusals and handovers have been testedMissing data, out-of-scope requests and unclear inputs all end with a person, not a guess.
Approval cannot be skippedNo tool, integration or retry path creates a record without passing the approval step.
Outputs describe themselves accuratelyThe agent says 'draft' when it created a draft and links to the record.
Duplicates are handledReruns and resubmissions find the existing record instead of creating another.
Ownership is namedA process owner approves future changes and an administrator applies them.
A regression set is savedThe same cases can be rerun after every quarterly Fusion update.

Questions and answers

Which security roles do agent builders and agent users need?

It depends on the product area and the agent. For the procurement Quote to Purchase Requisition Assistant, Oracle documents administrator duties for managing agents, a runtime duty for users who interact with agents on product pages, and extra privileges for the email account and the scheduled job⁶. Map the equivalent duties for your agent before testing, and if you secure your copy with a custom role, make sure the people and jobs that run it hold that role.

Can we use agent templates built by Oracle partners?

Yes, where your release includes them. Oracle's AI Agent Marketplace places partner-built agent templates alongside Oracle's prebuilt agents inside AI Agent Studio⁴. Treat a partner template like any other starting point: review its tools, credentials and instructions against your approval design before publishing it, and confirm who maintains it when Fusion is updated.

Can an agent team call systems outside Fusion?

Yes, through connector or REST-style tools and Model Context Protocol servers, with secrets held in AI Agent Studio's credential store⁴. Decide for each external call whether it reads or writes data. A write to an outside system, or back into Fusion through an integration, should pass the same approval step as a write made inside the application.

Should one agent team cover a whole department?

Rarely. A team built around one task with clear completion criteria is easier to test, secure and support. Departments usually end up with several small teams, each with its own process owner, tool scope and regression set.

Sources

Components of AI Agent Studio — Oracle · checked 10 October 2026
Create AI Agents Using Preconfigured Templates — Oracle · checked 10 October 2026
What's the difference between the Copy Template and Use Template options in AI Agent Studio? — Oracle · checked 10 October 2026
Oracle Expands AI Agent Studio for Fusion Applications with New Marketplace, LLMs, and Vast Partner Network — Oracle · checked 10 October 2026
Use the Approval Process Channel for Human Approval Nodes in AI Agent Studio (Fusion 26C) — Oracle · checked 10 October 2026
AI Agent: Quote to Purchase Requisition Assistant (Fusion 25D readiness) — Oracle · checked 10 October 2026

Plan your first Fusion agent team with us

Tell us the process, your Fusion release and the seeded agents you have looked at. We will suggest whether to extend a template, what the tool scope should be and where the approval step belongs.

Discuss an agent team

The parts of an agent team and what each controls

Copy a template, use a template or start from a blank team

Writing topics and instructions for an ERP task

Tool types and the scoping question for each

From blank canvas to published agent team

List what Oracle already provides

Create the team and name it for its job

Define agents and their roles

Attach topics and write instructions

Attach tools with least privilege

Add the approval step

Choose the model

Test against real exceptions

Secure, publish and monitor

A hypothetical quote-to-requisition assistant

Pre-release checks for a new agent team

Questions and answers

Sources

Related reading

More in Oracle ERP AI Agent Studio

Plan your first Fusion agent team with us