Oracle AI Agent Studio
Fusion's low-code builder for agents, workflow agents, agent teams, tools, topics, business objects and agentic apps. Included with Fusion, runs in each pod and inherits Fusion security.
Lesson63 terms. Each links to the lesson that teaches it.
Fusion's low-code builder for agents, workflow agents, agent teams, tools, topics, business objects and agentic apps. Included with Fusion, runs in each pod and inherits Fusion security.
LessonOracle's pro-code tooling: a CLI skill that gives coding assistants Fusion context, plus a VS Code extension. Artifacts live as files in Git; fetch, validate, test and save drafts from code.
LessonAgentic Testing and Lifecycle Automation Suite in Fusion AI Studio. Records live runs and replays them as workflow, conversation and app tests with path assertions and an LLM judge.
LessonWhere partners publish agents and agentic apps for Fusion customers, approved by Oracle per quarterly release.
Agents, workflows, policies and apps Oracle ships with ORA_ codes. Duplicate and extend them; never edit in place, so quarterly updates still apply.
Each environment (dev, test, prod) is a pod. Artifacts move by export, import and publish, and imports carry no secrets.
LessonAn agent built as a graph of typed nodes, saved as a .wf file. Deterministic steps do the volume and rules; model steps do the judgement. The default unit of build.
A supervisor classifies the request and routes it to worker agents, typically a read-only query agent and a separate execution agent that holds the write tools.
LessonOne workflow step: AGENT, LLM, CODE, CONDITION, SWITCH, PARALLEL, LOOP, WHILE, WAIT, HUMAN, POLICY, WORKFLOW and tool nodes. Display names must be unique.
LessonA model that reasons over instructions, topics and tool schemas, calls tools and loops (up to maxInteractions, 1 to 20) until it can answer. Use it only when tools are attached.
One prompt in, one answer out, no tools. Cheaper and more predictable than an AGENT node: summaries, classification, drafting.
LessonSandboxed JavaScript that filters, joins, calculates and caps rows. Reports 0 tokens; output lands at $output.result. The most-used node type in production workflows.
Deterministic routing. CONDITION has true and false outcomes; SWITCH has one outcome per case, and its fallback branch is labelled default.
Everything a node can read: $nodes (other nodes' outputs), $variables, $system, $app, $triggers, $user, $error. Written as {{$context.$nodes.X.$output}} in inputs and prompts.
Runs another published workflow as a step. A failed child leaves no output; a child that pauses on HUMAN pauses the parent.
LessonThe node a run jumps to when this node fails. The handler reads $context.$error.$errorMessage. Failure is never an edge.
What starts a workflow: REST (chat, apps, API), SCHEDULE (as a service account), EMAIL (an inbound mailbox), EVENTS/signals, or CAPABILITY (Fusion Claw).
LessonEvents one workflow publishes and others subscribe to, with retries and backoff. Fusion business events still arrive through Oracle Integration.
LessonA reusable set of business rules and instructions carried into an agent's context on every call.
A policy is generated from an SOP document: an LLM writes deterministic JavaScript functions with the rules as literals, plus test cases. Templates lock names and return shapes.
LessonRuns a policy function with no LLM call: same input, same decision. Oracle's AP Payment Exceptions Advisor uses 12 of them.
LessonPauses a run for a person over CHAT, EMAIL or APPROVAL_PROCESS (26C). Its only edge is success, and a timeout takes it too, with $output = TIMEDOUT.
A Fusion approval definition AI Studio workflows can route to: rules on payload attributes, approvers by user, management hierarchy or representative, notified by email, SMS, push or bell.
LessonA graph pattern: CODE classifies the reply and a CONDITION lets only a real confirmation reach the execute agent. A prompt-only "confirm first" failed 3 runs in 4.
LessonA capability an agent can call, described by a schema the model reads. Every attached tool's schema rides in context on every call.
LessonA Fusion REST resource registered in AI Agent Studio, such as purchaseOrders, workOrders or invoices, with fields, child resources and actions.
LessonReads and writes business objects over Fusion REST as the signed-in user, so Fusion roles and data security decide the rows.
LessonCalls a service outside the pod over HTTPS: Oracle Integration, ORDS, OTM, third parties. Runs as the stored credential, not the user.
LessonConnects to an MCP server, lists its tools (tools/list) and calls one (tools/call). One MCP tool brings the whole server's tool set into context.
Retrieves relevant chunks of uploaded playbooks and policies, so answers follow your own rules.
LessonContent connectors bring SharePoint files and crawled sites into retrieval; client connectors put agents into Teams and Slack.
Email, deep link, user session, calculator and web search tools, plus the seeded multi-file processor for attachments.
A Fusion workspace where several workflow agents share context around one process: panels, priority actions, a summary, communications and chat.
LessonA tile backed by a workflow agent and rendered at the InitDisplay stage, from <oraInfoDisplay> widget blocks. Bare JSON shows as raw text.
Why the app called the workflow, in $context.$app.$OraMessageHint: InitDisplay, InitActions, Summary, Query, InvokeAction, InitCommunications and more.
Ranked <oraInsight> actions built at InitActions from panels that set includeInActions; clicking one runs ora.Invoke("actionN", …).
Chat inside the app. The Ask box and every action click go to the app's single queryAgent, running its published version; a panel's own chat goes to that panel's workflow.
LessonRuntime values in $OraAppContext, such as a plant or business unit, that every panel reads, so users can switch organisation without a second app.
The older way to call an agent from outside Fusion: POST invokeAsync returns a job id, then poll its status until COMPLETE or ERROR. Oracle is replacing it with A2A, so build new callers on A2A.
Oracle's outcome engine inside agentic apps. A frontier model plans from the app's goals and capabilities, deterministic compute executes across Fusion data, and it re-plans until the outcome is reached.
LessonA goal set in the app's Priorities panel with a mode. Research returns a plan and changes nothing; Full Auto executes, inside your approvals. Outcomes can repeat.
LessonSet in App Settings once autonomous outcomes are enabled: goals, skills, capabilities, additional agents and example outcomes.
LessonA workflow agent declared as something Claw may plan with, under Settings › Triggers › Capabilities in the workflow builder.
LessonThe envelope sets the data, processes, roles, permissions and limits; the Trust Harness checks every proposed action against it before it touches Fusion.
LessonThe audit record of every Claw run: what was asked, what was done, which actions were governed and what was handed off.
LessonOne layer that understands a user's goal and selects the right Fusion capability: an agentic app with Claw, an agent, or a business action. Context and authority travel with the work.
LessonThe Fusion-native front door to Companion, with Work, Research and Full Auto modes, and # to insert business objects.
LessonEvery interactive call carries the running user's Fusion token, so the agent is bound by that user's roles and data security. Fusion makes the final authorisation decision.
LessonThe Credentials tab: bring-your-own model keys, email accounts (inbound or approver) and storage for chat uploads. REST and MCP auth is set on each tool.
LessonScans of requests and responses for prompt injection, toxic content and personal data, inside Oracle's defense-in-depth.
LessonMonitoring (or Metrics), Evaluation, Tracing, Reporting and Observability: the Monitoring and Evaluation tab. Errors, latency, tokens, AI Units and a trace of every run.
LessonScores answers from an evaluation set against reference answers for correctness, groundedness and relevance, with a score and an explanation.
LessonOracle's meter. Premium models: 5 AU per started 10K tokens per node. Basic: 0. A generated document: 10 or 20. 1 AU lists at USD 0.01.
LessonPremium: OpenAI models on a zero-retention endpoint outside Oracle. Basic: GPT-OSS-120B on OCI, data stays in Oracle Cloud. BYO: your key and endpoint, metered at multipliers.
LessonAgents built in an Oracle Integration project's AI tab, using registered integrations as their tools. Arrived in OIC 25.10 with projects as MCP servers and human-in-the-loop.
LessonAn Oracle Integration project exposed at /mcp-server/v1/projects/<project>/mcp: tools/list returns its registered integrations, tools/call runs one, as the OAuth client rather than the user.
An open protocol for agents to discover and call each other: an agent card describes an agent, and work moves as tasks and messages. Fusion exposes published agents through what Oracle calls a pragmatic subset of it, for other agents and for apps and integrations, as it replaces invokeAsync.
Fusion Setup and Configuration Registry: a public Oracle repository of 1,124 skills that let a coding agent map requirements to setup tasks, build a run book and load setup data after approval. Sample code, not a supported product.
LessonOracle's implementation platform (formerly Cloud Success Navigator). Its agents cover discovery, design, configuration (built on FSCR), testing, integration and innovation.
LessonFrom 8 Oct 2026, customers without a Visual Builder instance can no longer provision one. Existing instances continue, with no end-of-life date; Visual Builder Studio for extending Fusion pages is unaffected.
LessonIn APEX 26.1 (GA May 2026), AI Agents with AI Tools and tool guardrails, natural-language page creation and APEXlang application files. APEX connects to a model you choose; it ships none.
LessonAn open lakehouse with a governed catalog (supporting A2A and MCP), an Agent Hub and one workbench for data and AI teams. Fusion Data Intelligence data is available in it.
LessonOracle AI Database 26ai's no-code agent builder that runs as a container next to your data, in the cloud or on-premises, with prebuilt data agents, MCP tools and A2A.
LessonIndependent course, not an Oracle product. Based on Oracle documentation, release notes and hands-on builds; check Oracle's documentation for your release.