GLOSSARY

What is an AI Agent?

Table of content
Definition

Definition

An AI agent is a software system that uses an AI model to interpret a goal, plan or choose actions, use tools or data, observe results, and continue working toward an outcome.

TL;DR

  • An AI agent can take multiple steps toward a goal instead of producing only one response.
  • Agents may plan, call tools, retrieve information, inspect results, and revise actions.
  • The model is one part of the agent; tools, memory, context, and guardrails also matter.
  • Agents are useful when work requires decisions across several steps or systems.
  • Reliability depends on clear goals, permissions, evaluation, and high-quality context.

What makes something an AI agent?

Suppose a system is asked to review an existing onboarding flow and propose improvements.

A simple generative model might produce suggestions from one prompt.

An agent could inspect existing screens, read a PRD, analyze analytics, map the flow, identify missing states, generate a prototype, compare the result with the design system, and revise it before returning the final artifact.

The key difference is the multi-step loop between reasoning, action, observation, and adjustment.

Common parts of an AI agent

  • model
  • goal or instruction
  • tools
  • retrieval and context
  • memory or state
  • permissions
  • evaluation and guardrails

AI agent vs. chatbot

A chatbot primarily conducts a conversation. An agent can take actions outside the conversation through tools and can maintain a multi-step objective.

The same product can include both conversational and agentic behavior.

AI agent vs. generative AI

Generative AI is a capability for creating outputs. An agent wraps that capability in a process that can decide what to do next.

Why do context and permissions matter?

An agent acting on an existing product needs the right product context and limited tool permissions. Otherwise it can make locally plausible decisions that violate product rules or perform unintended actions.

Common mistakes

Calling every AI feature an agent

One-shot generation is not necessarily agentic.

Giving broad tool access without guardrails

Actions should match the task and user authorization.

Ignoring evaluation

Agents need checks for whether intermediate and final results are acceptable.

The bottom line

An AI agent asks: what sequence of information gathering and actions should the system take to accomplish this goal, and how should it adapt after each result?

Related terms

Generative AI · Context engineering · Product context · Multimodal AI

Relevant Figr resource

Read AI for Product Design for agentic design workflows grounded in existing products.

Related Figr Projects

No items found.