An Agent is a system that uses an LLM to make decisions and take actions to achieve a goal.
An Agent adds things around the LLM:
User → Agent → LLM → Action → Observation → LLM → ...
Imagine an automated workflow, Agent is really look like a flow with a big different thing: Agents have a non-deterministic flow but automations have a deterministic flow which is why Agents useful when we don't have a predfined path.
An Agent can look like an automated workflow, but there is one key difference: who decides what happens next?
In a workflow, the developer defines the execution path:
File → Extract → Count Words → If > 300 → Good Article
In an Agent, the developer defines the goal, while the Agent dynamically decides the next action based on its goal, state, and observations:
Goal → Decide → Act → Observe → Decide → ...
So:
- Workflow: Developer controls the flow.
- Agent: Agent dynamically controls the flow.
This makes Agents useful when the path to the goal cannot be fully predefined.
There is no single universal definition, but an Agent usually has these characteristics:
- Goal — something the Agent is trying to achieve, such as creating a landing page.
- Decision Making — it can decide what action to take based on its goal, current state, and observations. Some decisions can be deterministic, while others can be LLM-based.
- Tools — capabilities the Agent can invoke, such as external systems, internal functions, APIs, or MCP tools.
- State — the current state of the Agent during execution. This is not simply the history of actions.
- Observation — the information or result the Agent receives after taking an action and uses to decide what to do next.
- Iteration — it can perform multiple steps instead of generating a single response. This allows the Agent to evaluate the outcome of its actions and decide whether to continue or finish.
For example, consider a simple customer support Agent.
The user says:
"Where is my order?"
The Agent might:
User Request → Understand Goal → Decide Action → Call Order API → Observe Result → Decide What to Do Next
If the information is sufficient, the Agent finishes:
Decide → Respond
If it is not, the Agent can continue:
Decide → Take Another Action → Observe → Decide → ...
The important part is that the Agent is not simply generating an answer. It is deciding, acting, observing, and iterating based on its goal, current state, and available information.