AI agent
In brief
An AI agent is a system that uses a language model to pursue a goal through several steps: reasoning, using tools, observing the result and adjusting.
Sources[1]
Definition
An AI agent is a system in which a language model does more than answer once: it decides on actions (searching the web, reading a file, calling an API, running code), observes their result, then continues until it reaches the goal or stops.
Typical components
- Language model: the “brain” that reasons and chooses actions.
- Tools: functions the agent can call to act or gather information.
- Execution loop: alternating between reasoning, acting and observing. The ReAct approach (Yao et al.) formalised this combination.
- Memory: keeping the task context, sometimes longer-term memories.
- Safeguards: limits on actions, human approval for sensitive operations.
Benefits
Automating multi-step tasks (document research, data processing, software development) that previously required continuous human involvement.
Limits and risks
- Compounding errors: a mistake at one step can carry over to the next ones.
- Actions with real effects: an agent connected to tools can send, modify or delete data; its permissions must be limited and sensitive actions approved.
- Cost and predictability: several model calls per task, and behaviour that is sometimes hard to anticipate.
- Security: external content read by the agent may contain malicious instructions (prompt injection).
Spotted a mistake or have an idea about this entry? Let us know