Explain what an AI agent is — a model that calls tools and acts step by step — and describe the reason–act–observe loop with a human in the loop.
From chatbot to agent: tools and the ReAct loop
AI agents and tool use
A plain LLM can only do one thing: generate text. It cannot browse the web, run calculations, check a live database, or send an email — at least not on its own. Tool use (also called function calling) bridges this gap. The model is given a set of tools — search, a code interpreter, an API call, a calendar — and it can decide to invoke one, receive the result, and continue reasoning. Instead of just talking about what it would do, the model can actually do things.
An AI agent is a model that uses tool calling to act across multiple steps toward a goal. Rather than answering in a single response, the agent plans, acts, observes the outcome, and then reasons about what to do next. This cycle is often called the ReAct loop: Reason → Act → Observe → Reason again. For example, asked to summarize all customer complaints filed this week, an agent might: reason that it needs to query the database; act by calling a database API; observe the returned list; then reason that some entries are duplicates; act to filter them; and finally generate the summary.
The key distinction from a plain chatbot: a chatbot produces text in one shot. An agent pursues a multi-step plan and uses real tools to interact with systems outside the model.
Humans stay in the loop for oversight. Agents can make mistakes, misinterpret a goal, or take unintended actions — especially when tools can change data or send communications. Good agent design includes checkpoints where a human reviews and approves the next action before it is taken, particularly for irreversible or high-stakes steps.
Lesson notes
AI agents and tool use
A plain LLM can only do one thing: generate text. It cannot browse the web, run calculations, check a live database, or send an email — at least not on its own. Tool use (also called function calling) bridges this gap. The model is given a set of tools — search, a code interpreter, an API call, a calendar — and it can decide to invoke one, receive the result, and continue reasoning. Instead of just talking about what it would do, the model can actually do things.
An AI agent is a model that uses tool calling to act across multiple steps toward a goal. Rather than answering in a single response, the agent plans, acts, observes the outcome, and then reasons about what to do next. This cycle is often called the ReAct loop: Reason → Act → Observe → Reason again. For example, asked to summarize all customer complaints filed this week, an agent might: reason that it needs to query the database; act by calling a database API; observe the returned list; then reason that some entries are duplicates; act to filter them; and finally generate the summary.
The key distinction from a plain chatbot: a chatbot produces text in one shot. An agent pursues a multi-step plan and uses real tools to interact with systems outside the model.
Humans stay in the loop for oversight. Agents can make mistakes, misinterpret a goal, or take unintended actions — especially when tools can change data or send communications. Good agent design includes checkpoints where a human reviews and approves the next action before it is taken, particularly for irreversible or high-stakes steps.