Skip to main content

Core Concepts

Before building an agent in AI Agent Studio, it helps to know the four building blocks you'll be working with.

AI Agent​

The top-level object you build: a goal-driven piece of software with a defined purpose (for example, "resolve level-1 customer support tickets" or "review expense reports for policy violations"). An agent is composed of one or more skills and is powered by the Process Reasoning Engine.

Skill​

A skill is a reusable capability you give an agent — a self-contained unit of work such as "categorize inquiry priority" or "update the inventory system." Skills are what make agents modular: the same skill (say, "look up customer account") can be reused across multiple agents. Skills commonly leverage prompts — optimized instructions/templates that tell the underlying model exactly how to complete that particular task according to defined guidelines (for example, a prompt that checks whether a request meets policy criteria for escalation to a specialist team).

Process Reasoning Engine (PRE)​

The PRE is the reasoning layer underneath every agent. It's what lets an agent look at a specific case, decide which skill(s) to invoke and in what order, call out to bots, APIs, or documents as needed, involve a human when appropriate, and adjust its plan as new information comes in — instead of following one fixed sequence every time.

AI Guardrails​

Guardrails are policies enforced around the agent, not inside your prompts. They block unsafe or non-compliant actions before they happen, so an agent can be given real autonomy without giving up governance. When you're designing an agent, it's worth deciding early which actions should always require guardrail approval versus which are safe to let the agent take independently.

How these fit together​

A typical agent definition looks like: goal → one or more skills the agent can use to pursue that goal → guardrails constraining what it's allowed to do → the PRE deciding, case by case, which skills to call and when a human needs to step in.

Next: Build your first agent walks through creating a simple agent end to end.