An AI agent is a system that uses an LLM to decide, within defined constraints, which actions to take toward a goal — typically by selecting and calling tools, observing results, and deciding what to do next. Exact definitions vary somewhat across frameworks and papers, but this decision-loop behavior is the common thread.
The Minimum Ingredients of an Agent
| Component | Role |
|---|---|
| A goal or task | What the agent is trying to accomplish |
| An LLM | Makes decisions about what to do next, given the current state |
| Tools/actions | The concrete things the agent can actually do — search, query a database, call an API |
| Observation | Reading back the result of an action to inform the next decision |
| A stopping condition | Knowing when the goal is met, or when to give up/escalate |
Remove the tools and the observation loop, and you just have a chatbot — see Agent vs Chatbot.
Simplified Example (Conceptual, Not a Specific Framework)
goal = "Find today's USD to INR exchange rate and convert 500 USD"
while not done:
decision = llm.decide_next_action(goal, history)
if decision.action == "call_tool":
result = run_tool(decision.tool_name, decision.arguments) # validated first
history.append(result)
elif decision.action == "final_answer":
return decision.answerThis is the essential shape of an agent loop — see Agent Loop for the fuller version, including validation and error handling.
Practical Use Case
A data-analyst agent given access to a SQL-query tool: it can decide to run a query, look at the result, realize it needs a follow-up query to answer the full question, run that too, and only then produce a final answer — a task that a single fixed prompt-response can't handle because the exact queries needed depend on what the first query returns.
Common Mistakes
- Treating "has an LLM in it somewhere" as sufficient to call something an agent — the decision loop and tool use are the defining traits, not just LLM usage
- Assuming an agent always needs multiple tools or complex logic — a single-tool agent that can decide whether to call that tool is still agentic, as opposed to a system that always calls it unconditionally (that's closer to a fixed workflow)
Interview Relevance
"Define an AI agent in one sentence" is a common opener — the strongest answers mention the LLM-driven decision loop and tool/action selection specifically, not just "AI that does things automatically."
Practice Question
List the five minimum ingredients of an agent (from the table above) for a system designed to book a meeting by checking a calendar and sending an invite.
