Oracle AI Agent Studio Bring Your Own LLM
Oracle AI Agent Studio lets you choose the large language model that powers your Fusion agents. You can use an Oracle-provided model or add your own model by supplying credentials for an external provider. Per Oracle's documentation, the provided options include OpenAI (accessed via Azure), Google Gemini, and Anthropic. This page explains what the Oracle AI Agent Studio LLM choice actually controls, how each option is configured, and why the model you pick changes how your agents behave — and therefore how they must be tested.
Agent Studio is generally available in Fusion (GA since release 25C). The model-selection and add-your-own-LLM capability is documented by Oracle; where a specific detail is release-dependent, confirm the current behaviour against the Oracle documentation linked at the foot of this page rather than assuming it.
Last reviewed: 19 July 2026
Scope note. This page covers one narrow topic: the LLM behind an agent — Oracle-provided models, adding your own model via credentials, and the configuration, security, cost, and testing implications of that choice. For the broad overview of what Agent Studio is and everything it can do, see the Oracle AI Agent Studio page. For how the platform is put together internally, see AI Agent Studio architecture. This page does not re-explain agents, topics, or tools.
What "Bring Your Own LLM" Means in Agent Studio
Every agent you build in Agent Studio calls a large language model to interpret instructions, reason over a topic, decide which tool to invoke, and generate a response. By default an agent uses a model Oracle makes available inside Fusion AI. "Bring your own LLM" (BYO LLM) means you are not limited to that default: Agent Studio lets you register an external model by supplying the credentials for it, and then point agents at that model instead.
Be precise about what this feature is and is not. It is a way to select the model an agent uses. It is not a way to fine-tune Oracle's data on a third-party model, and it does not change the rest of the agent — the instructions, topics, tools, knowledge sources, and approval gates stay the same regardless of which model sits underneath.
Two supply paths exist. With an Oracle-provided model, Oracle has already established the integration and you simply choose it. With your own model, you provide the provider credentials so Agent Studio can call that endpoint on your behalf. Both paths make the model available to agents; they differ in who owns the model relationship, where the request goes, and what you are responsible for. The next section lays out the options side by side. Availability: model selection is documented for Agent Studio (GA since 25C); confirm the current provider list and any regional constraints with Oracle.
Oracle-Provided Models vs Your Own LLM
Per Oracle's add-your-LLM documentation, Oracle-provided options include OpenAI (via Azure), Google Gemini, and Anthropic; you can also add your own model by supplying credentials. Options are given at the provider level — this page does not state specific model version numbers, which change and should be read from current Oracle and provider documentation.
| Option | How it's added | Considerations |
|---|---|---|
| Oracle-provided: OpenAI (via Azure) | Selected from the Oracle-provided model options; Oracle establishes the integration (OpenAI accessed through Azure). | Least setup on your side. The request path runs through Oracle's provided integration; review where prompts and data are processed and any regional terms with Oracle. |
| Oracle-provided: Google Gemini | Selected from the Oracle-provided model options; integration provided by Oracle. | No provider account to manage yourself. Confirm data-handling and residency behaviour for the Gemini option with Oracle before production use. |
| Oracle-provided: Anthropic | Selected from the Oracle-provided model options; integration provided by Oracle. | No provider account to manage yourself. As with the others, verify the processing path and terms rather than assuming they match another provider. |
| Your own LLM (via credentials) | You register the external model in Agent Studio by supplying provider credentials so it can call that endpoint. | You own the provider relationship, the account, its cost, and its data terms. Credential management, endpoint availability, and the provider's data-residency guarantees become your responsibility. |
Source: Oracle "Add Your LLM" documentation (linked in Official Oracle references). Provider list per Oracle; individual model versions are not stated here because they change.
Configuring the Model for an Agent
At a high level, using an Oracle-provided model is a selection step, while adding your own model is a registration step followed by that same selection. Exact screen paths and field labels are release-dependent, so the checklist below is generic and framed around the decisions rather than specific navigation — confirm the current steps in the Oracle documentation.
Model configuration checklist
- 01Decide the supply path. Choose an Oracle-provided model for the lightest setup, or your own model where you need a specific provider, account, or contractual terms.
- 02For your own model, obtain and store credentials. Register the external model in Agent Studio using the provider credentials, and manage those secrets under your organisation's key-handling standards.
- 03Assign the model to the agent. Point the agent at the chosen model so its reasoning and generation run on that provider.
- 04Review the data path. Confirm where prompts, business context, and responses are processed for that provider before any production data flows through the agent.
- 05Validate behaviour. Because the model changes how instructions are interpreted, exercise the agent's prompts and instructions and its tool calls against the new model before trusting it.
- 06Record the choice. Note which model each agent uses so a later model change is a controlled, auditable event rather than an untracked one.
The mechanics of registration are handled inside Agent Studio; how those pieces fit the wider platform is covered on the AI Agent Studio architecture page.
Security & Data-Residency Considerations
The model choice determines where your prompts and the business context an agent assembles are sent for processing, which makes model selection a data-governance decision, not just a quality one. The considerations below are the questions to resolve with Oracle and, for your own model, with your provider.
| Consideration | Why it matters | What to confirm |
|---|---|---|
| Where prompts are processed | Agent prompts can carry business data assembled from Fusion context. | The processing location and region for each Oracle-provided or own model. |
| Data residency | Regulatory and contractual obligations may restrict where data may leave. | Whether the option meets your residency requirements for the data involved. |
| Data retention & training use | Whether inputs are retained or used to improve a model varies by provider. | The provider's retention and training terms for the exact option chosen. |
| Credential handling (own model) | Your provider credentials are stored so Agent Studio can call the endpoint. | How secrets are stored, rotated, and scoped under your standards. |
| Access & approval gates | A model change can alter what an agent proposes to do. | That human-approval and role controls still apply after a model change. |
| Auditability | You need to know which model produced a given agent action. | That the model in use is recorded and reviewable per agent. |
Agents continue to operate within existing role-based permissions and approval frameworks, but a model swap changes the reasoning that leads to an action — so the controls around that action deserve a fresh look each time the model changes.
Cost Considerations
Model choice has a cost dimension, but the specifics depend on your Oracle agreement and, for your own model, your provider contract. This page does not quote prices because they change and vary by customer. Treat the points below as the axes to price out, not as figures.
- •Provided vs own. An Oracle-provided model folds the model relationship into your Oracle arrangement; your own model is billed by your provider under your account.
- •Usage drivers. Cost generally scales with how much the agent is used and how much context each call carries — more tools, longer instructions, and larger knowledge context mean more work per request.
- •Model tier. Providers offer different model tiers at different rates; a more capable model may cost more per call but need fewer retries or less correction.
- •Operational overhead. Bringing your own model adds account, credential, and monitoring responsibilities that carry their own cost in effort.
Confirm the actual commercial terms for each option with Oracle and your provider before committing an agent fleet to a given model.
Why a Model Change Means Re-Testing the Agent
The single most practical consequence of BYO LLM is this: the model is the reasoning engine of the agent, so changing it can change the agent's behaviour even when every instruction, topic, tool, and knowledge source is untouched. A model may interpret the same prompt differently, choose a different tool, phrase a response differently, or reach a different conclusion at a decision point. None of that is visible from configuration alone — it only shows up when you run the agent.
Because of that, treat a model swap as a change that requires validation, not a like-for-like substitution. The discipline you apply when Oracle ships a quarterly update applies here too: establish how the agent behaves today, make the change, then re-check that instructions are still followed, tool selection is still correct, outputs are still accurate and safe, and approval gates still fire.
This is exactly the ground covered by agent evaluation and testing — defining what "good" looks like for an agent and measuring against it. A model change is one of the clearest triggers for running that evaluation again. Reviewing how the agent's prompts and instructions perform under the new model is a sensible first step, because instruction-following is often where model differences surface first.
How SyntraFlow Helps
SyntraFlow does not choose your model or run your Oracle AI agents. Where it can help is around the change: giving you a way to see whether an agent still behaves as expected after you switch the model underneath it.
Baseline & compare
SyntraFlow can be configured to capture how an agent responds today and compare it after a model change, so behavioural differences surface rather than hide.
Model-change regression
SyntraFlow helps organisations assess whether instruction-following, tool selection, and approval behaviour hold up when the LLM behind an agent is swapped.
Config-aware change tracking
SyntraFlow can connect configuration insight with test planning so a model change is treated as a tracked event with a matching validation scope.
A note on scope. These are complementary capabilities framed at the level of validating agent behaviour around a model change; exact coverage for your agents is confirmed at assessment rather than assumed here. To compare model and agent configuration across environments before you promote a change, SyntraFlow's Configuration Intelligence can highlight where the setup differs — so a validated agent reflects the configuration you intend, not a coincidental match.
Official Oracle References
The facts on this page are drawn from Oracle's own documentation. Confirm current behaviour, provider lists, and model versions against these sources:
- •Add Your LLM — Oracle Fusion AI documentation
- •AI Agent Studio key capabilities — Oracle Fusion AI documentation
- •Oracle Fusion AI — product overview
- •AI Agent Studio 25C readiness (GA)
Last reviewed: 19 July 2026
Frequently Asked Questions
Can I use my own LLM in Oracle AI Agent Studio?
▼
Yes. Oracle's documentation describes adding your own LLM to Agent Studio by supplying the credentials for an external provider, and then assigning that model to your agents. You can also use an Oracle-provided model without managing a provider account yourself.
Which models does Oracle provide for Agent Studio?
▼
Per Oracle's add-your-LLM documentation, the Oracle-provided options include OpenAI (accessed via Azure), Google Gemini, and Anthropic. Specific model versions change over time, so read the current version from Oracle and the provider documentation rather than relying on a fixed number.
What is the difference between an Oracle-provided model and my own model?
▼
With an Oracle-provided model, Oracle has established the integration and you select it. With your own model, you register an external endpoint using your provider credentials and own the account, its cost, and its data terms. Both make the model available to agents; they differ in who owns the model relationship and what you are responsible for.
Where does my agent data go when it calls the model?
▼
The prompt an agent sends — which can include business context — is processed by the model provider you selected. Because this varies by option, confirm the processing location, data residency, retention, and training-use terms with Oracle and, for your own model, with your provider before production data flows through the agent.
Does changing the model change how my agent behaves?
▼
It can. The model is the reasoning engine of the agent, so a different model may interpret the same instruction differently, select a different tool, or produce a different response — even when the instructions, topics, and tools are unchanged. That is why a model change should be validated rather than treated as a like-for-like swap.
Should I re-test an agent after switching its LLM?
▼
Yes. A model change is a clear trigger for re-running agent evaluation: check that instructions are still followed, tool selection is still correct, outputs are still accurate and safe, and approval gates still fire. See the agent evaluation and testing page for how to define and measure that.
How much does bringing your own LLM cost?
▼
That depends on your Oracle agreement and, for your own model, your provider contract. An Oracle-provided model folds into your Oracle arrangement; your own model is billed by your provider. Cost generally scales with usage and the amount of context each call carries. Confirm the actual commercial terms with Oracle and your provider.
Is Agent Studio's model selection generally available?
▼
Oracle AI Agent Studio is generally available in Fusion (GA since release 25C), and model selection including adding your own LLM is documented by Oracle. Where a specific detail is release-dependent, confirm current availability and any regional constraints with Oracle rather than assuming it.
Planning a Model Choice for Your Oracle AI Agents?
Understand the security, data-residency, cost, and testing implications of your Oracle AI Agent Studio LLM choice — and how to validate agents when the model changes. Talk to SyntraFlow about an assessment scoped to your agents.