Chapter 5: No-code Select AI Agent

Canvas

Flow name: SELECT AI - no PL/SQL. All objects get a C3_ prefix so they don't collide with the PL/SQL-created objects from Chapter 3.

Every node is in + Create new mode, and there are two kinds of wires:

• Definition chain (dark ports) shows how the team is assembled: In-Database Tool → Task Choose AI Tools · Task Task → Agent Task · Agent Agent → Team Agents
• Message flow (blue ports) shows what happens at runtime: Chat input Message → Team Prompt · Team Result → Chat output Message

In-Database Tool ──► In-Database Task ──► In-Database Agent ──► In-Database Team ──► Chat output
                                                                        ▲
                                                           Chat input ──┘

In Create new mode the Task node gets a Choose AI Tools input and the Team node an Agents input. That is how the definition chain is wired.

Node settings

In-Database Tool

Instruction:

Use this tool for any question about sales, revenue, profit, discounts, units, orders, customers, products, geography or time periods. It translates the question into SQL over the sales database and returns the result.

In-Database Task

The instruction is the task instruction from Chapter 3, with two changes: it names the C3_SALES_SQL tool, and it does not ask for the runsql action. {query} is replaced with the user's question at runtime.

In-Database Agent

The Agent Role is the role from Chapter 3 plus one sentence that tells the agent how to refuse off-topic questions:

You are a sales analyst for a consumer electronics company. You answer business questions using data from the sales database only, never from general knowledge. If a question is not about sales data, reply that you can only answer questions about the sales database. Revenue is gross unless the user asks for net revenue (revenue minus discount). Profit is revenue minus fixed and variable cost. Include all order statuses unless the user explicitly asks about a specific status. Sales data covers January 2024 to September 2026.

In-Database Team

The forms have no human-tool setting. The agent never paused for clarification in my tests, so the defaults work for this non-interactive use.

Tests

Checking the database

The canvas does not keep its own copy of the team: in Create new mode OPAF creates real Select AI Agent objects in the database, as the user of the database connection. As OABOOTCAMP:

SELECT agent_team_name, status FROM user_ai_agent_teams;

AGENT_TEAM_NAME       STATUS  
--------------------- ------- 
OABOOTCAMP_SALES_TEAM ENABLED 
C3_SALES_TEAM         ENABLED

Both teams are there: OABOOTCAMP_SALES_TEAM, created in PL/SQL in Chapter 3, and C3_SALES_TEAM, created from the OPAF canvas.

The test answers above are the same as those of the PL/SQL-built team in C2. The objects are identical whether I create them with DBMS_CLOUD_AI_AGENT or on the OPAF canvas. The only prerequisite is a good Select AI profile over commented tables.

So far OPAF has reached the database over a database connection. In Chapter 6 I open the same agent to any MCP client with the Autonomous Database built-in MCP server.

Back to the introduction and list of chapters.