An AI agent is a program that can perceive its environment, reason about it, and act autonomously to achieve a defined goal. Where a plain model simply answers a question, an agent chains decisions, uses tools, and corrects its own course — much like a resourceful intern handed a mission rather than a single instruction.
Perceive, decide, act
At its heart an agent runs a perception–decision–action loop. It observes a state, picks an action, observes the outcome, then repeats until the goal is met. Formally, we seek a policy $\pi$ that maximizes the expected cumulative reward:
$$\pi^* = \arg\max_{\pi} \; \mathbb{E}\left[\sum_{t=0}^{T} \gamma^{t} \, r_t\right]$$
where $r_t$ is the reward at step $t$ and $\gamma \in [0,1]$ is the discount factor weighting the future.
Modern (LLM) AI agents
Recent agents use a large language model as their reasoning "brain," augmented with three essential building blocks:
- Tools: API calls, web search, code execution, calculators.
- Memory: retaining context across steps.
- Planning: breaking a complex goal into sub-tasks.
Agent vs. simple assistant
| Criterion | Assistant (chatbot) | AI Agent |
|---|---|---|
| Initiative | Reactive | Proactive |
| Steps | One reply | Multi-step |
| Tools | Rarely | Core of the system |
| Goal | Short-term | Full mission |
An agent does more than answer: it acts to reach a goal, one step at a time.