AI Agent

The four defining characteristics

An AI agent is defined less by its underlying model than by its behavior. Four characteristics are required for the label to fit:

1. Goal-oriented. You give the agent an objective (“respond to incoming support tickets”) rather than a step-by-step recipe.
2. Tool-using. The agent can call external systems (APIs, databases, files) to gather information or take action. Reading a knowledge base alone is not enough; that is just retrieval.
3. Autonomous within bounds. The agent decides what to do next on its own, within rules its operator has set. “Bounded autonomy” (described separately in this glossary) is the engineering pattern that makes this safe.
4. Stateful across steps. The agent maintains context as it works through a multi-step task, rather than starting from scratch every turn.

If any of these is missing, the system is something else, a chatbot, a workflow automation, a script. The combination is what makes an agent.

What an AI agent is not

• Not a chatbot. Chatbots respond to user input in a session; agents act autonomously, often without a user in the loop. See our detailed comparison.
• Not a Zapier automation. Traditional automation follows fixed if-this-then-that rules; agents make decisions about what to do next based on the situation, using natural language understanding.
• Not the underlying model. Claude or GPT-4 are large language models, agents are systems built on top of them with prompts, tools, and integrations.

How AI agents are built in practice

A production AI agent at theagency47 has four customizable components: a system prompt (the role and constraints), a knowledge base (documents the agent can retrieve from), tool integrations (APIs the agent can call), and an eval suite (test cases that verify behavior before launch). Each component is documented in this glossary as its own entry.

Typical build time for a single specialized agent: 14 days (Spark engagement). Typical cost: €2,500 to €15,000 for build, plus €20 to €500 per month per agent for running cost (full cost breakdown).

FAQ

Is “AI agent” just marketing language for chatbot?

No. The four characteristics above distinguish them architecturally. The marketing layer is muddier than the technical reality.

What is the difference between an AI agent and agentic AI?

”Agentic AI” is the broader pattern (AI systems that act, not just respond). “AI agent” is one instance of an agentic system. See the agentic AI entry.

Do I need an AI agent or just an AI tool?

If the work is conversational and bounded to one session, you probably need a chatbot or assistant. If the work runs in the background, involves multiple steps and multiple tools, you need an agent. Discovery call is the fastest way to find out.