Select AI Agent Concepts

Agentic Action

In the context of agents, an action is an instruction that triggers a tool with parameters. The platform runs the tool, handles errors, and returns an observation for the next reasoning step loop.

Actions connect planning to execution. They define which tool to run and with which inputs.

Agent

An agent performs tasks for a defined purpose. An agent is a configured worker that reasons about requests, selects tools, runs steps, evaluates results, and produces responses grounded in database context.

An agent encapsulates behavior, policies, and available tools for a particular job, such as returns processing or knowledge retrieval.

Agent Team

One or more agents that perform an agentic workflow. A team coordinates responsibilities, shares context, and runs agent-task pairs to complete multi-step interactions reliably.

Teams divide work among specialized agents and sequence their contributions. Teams keep shared context and produce a unified response.

Observation

The data returned from a tool run: rows, messages, or errors, that the agent records and evaluates during reflection. Observations guide the next reasoning step and provide evidence for the final response.

Task

A task represents a unit of work. It guides tool selection, parameter mapping, and execution policy, producing results that downstream steps can read and summarize.

Tasks specify the goal, inputs, tool choices, and guardrails. They return structured outputs that later steps can consume.

Tool

A tool performs an action such as updating data, retrieving documents, or calling external services. Tools take parameters, can run deterministically or non-deterministically, and return observations for reasoning.

Tools encapsulate repeatable operations. They keep side effects controlled and observable, supporting auditing and debugging.

In the Select AI Agent framework, tools represent functional building blocks that agents can use during task processing. A tool might be:

A deterministic tool, such as a PL/SQL function, which returns the same output for the same input.
A non-deterministic tool, such as a web search or LLM-based summarization, where outputs may vary depending on timing, network results, or randomness.

Agents select tools based on the task's needs and use the output (known as an observation) to reason, decide next steps, or generate responses. Tool metadata and run history are logged to support debugging, observability, and security.