Oracle details agent-builder updates for memory, model choice and deployment

Oracle is giving developers more control over what their AI agents remember, which models they use and how they run after deployment. An October 28, 2026, walkthrough details AI Data Platform updates spanning its visual flow builder and production settings, including conversation summaries, reusable skills, outside model connections and automatic compute scaling.

The changes did not all debut with that post. Oracle’s October 16 release notes already listed visual-builder skills, runtime identity, memory management, third-party models and autoscaling. The new walkthrough adds configuration details and describes deployment changes, including a way to update an agent at its existing endpoint without first taking the separate step of undeploying it.

Deciding what an agent keeps

Developers can enable automatic summarization when conversation history would exceed 80% of the selected model’s context window—the amount of material it can consider at once. The setting summarizes older messages rather than continuing to include the full history. Oracle says this helps retain important context while keeping the conversation within that limit.

That control is separate from long-term memory. In the Stored Memory tab, developers choose which model extracts memories and whether those memories carry across sessions started by the same user. By default, extracted memories stay within the session where they were created. Enabling summarization therefore is not the same decision as allowing an agent to remember someone across conversations.

A further setting controls what reaches the model during a task. The agent node’s Context tab lets developers select conversation history and tool outputs for inclusion. Oracle points to large SQL query results as one reason to use it: a database response can consume much of the space otherwise available for the conversation.

Outside models and reusable instructions

The visual flow builder now supports models from OpenAI, Google’s Gemini family and Anthropic, alongside models available through Oracle Cloud Infrastructure’s Generative AI service. Developers select a provider, save its API key in the Credential Store and configure the model for their agent. The external-provider configuration uses fields based on OpenAI software development kit parameters.

Reusable skills bring packaged task instructions and expertise into the same visual workflow. Oracle previously supported using skills through code; developers can now upload a local skill folder in an agent’s Skills tab or create and edit one there. The addition gives visual-builder users a way to reuse task-specific knowledge without leaving the builder to author each skill.

Scaling and deploying under a chosen identity

For production workloads, compute autoscaling can respond to API requests per second or average CPU utilization across the running compute replicas. Developers set minimum and maximum replica counts to bound capacity. In the testing playground, they can also change or deselect the hosting compute from its menu.

Deployment controls include an editable Agent2Agent card, which describes an agent’s capabilities and how to access it. Developers can publish the revised card with the deployed agent. The Run As setting separately chooses whether platform operations use the developer’s own identity or a service account, keeping deployment independent of an individual user’s access.

Using a service account requires registering its credentials and assigning the account’s user an appropriate platform role. The agent then operates with that account’s permissions. But Run As is not a single access switch: permission to invoke the agent and authentication to downstream tools remain separately configured.