Definition
An AI agent is a system that wraps a language model in a loop so it can pursue a goal over many steps rather than answering a single prompt. The model acts as the reasoning engine: it looks at the current situation, decides on an action, carries that action out using a tool, observes the result, and then decides what to do next — continuing until the task is complete or it gives up.
The perceive-reason-act loop
What separates an agent from a plain chatbot is the loop:
- Perceive — take in the goal plus any new information (a tool’s output, an error message, a user reply).
- Reason — the model thinks about what to do next, often planning several steps ahead.
- Act — it calls a tool or produces a result, and that output becomes the input to the next perception.
Because each step’s outcome is fed back in, the agent can recover from mistakes, gather missing information, and adapt its plan as it goes.
Tools, planning, and memory
Three ingredients turn a model into a capable agent:
- Tools let it act on the world — search the web, run code, query a database, or call an API. The model chooses the tool and its arguments.
- Planning lets it break a vague goal (“book me a trip”) into ordered sub-tasks and tackle them in sequence.
- Memory lets it carry context across steps, remembering what it has already tried so it does not loop forever.
Single call vs persistent agent
A single LLM call is a one-shot transaction: prompt in, completion out, no follow-through. An agent is goal-seeking: it persists across many calls, takes real actions, and judges its own progress. This distinction is why agentic systems can complete multi-step workflows — but also why they are harder to control and secure, since each autonomous action carries risk.
Why it matters
Agents are the frontier of practical AI deployment, powering coding assistants that edit and run code, research tools that browse and synthesise sources, and automation that operates software on a user’s behalf. Understanding the perceive-reason-act loop clarifies both the power of these systems and the new safety challenges — like prompt injection through tool outputs — that come with giving a model the ability to act.