What is an AI Agent?

Overview

An AI agent is a system where an LLM decides what to do next in pursuit of a goal. A chatbot answers one message at a time. An agent takes a goal such as "find the three cheapest flights and put them in a spreadsheet", then loops through thinking, calling tools, observing the results, and deciding the next step until the goal is met or it gives up.

An agent has four parts: a model that reasons, tools it can call (APIs, search, code execution, databases), memory of what has happened so far, and a control loop with limits. The engineering challenge is less about intelligence and more about reliability: bounding cost and steps, validating tool inputs, handling failures, asking a human before risky actions, and tracing every decision.

When to use it

• Tasks that need several steps whose order is not known in advance.
• Workflows that combine reading information and taking actions across systems.
• When a fixed pipeline cannot cover the variety of user requests.

Where it shows up in interviews

Where it is used in real software

Key terms

Tool

A function the agent can call, described by a name, purpose, and input schema.

Agent loop

The repeated cycle of reason, act, observe until done.

Memory

Short-term conversation state plus optional long-term stored facts.

Guardrail

A rule or check that limits what the agent may do.

Human in the loop

Requiring human approval before certain actions.

How it works, step by step

1. 1
   Receive a goal

   The user states an outcome, along with context such as their identity and permissions.

2. 2
   Reason about the next step

   The model reviews the goal, memory, and available tools, then picks an action.

3. 3
   Call a tool

   The runtime validates the arguments and executes the tool with the user's permissions.

4. 4
   Observe and update memory

   The tool result is added to the conversation so the model can use it.

5. 5
   Repeat or finish

   Loop until the model returns a final answer, a step limit is hit, or a human must approve.

Chatbot vs workflow vs agent

Three ways to handle "Refund my order 4821".

Use a fixed workflow when steps are known; use an agent when requests vary widely.

Implementation

1import OpenAI from "openai";2const openai = new OpenAI();3 4const tools = {5  getOrder: async ({ id }: { id: string }) => orders.find(id),6  issueRefund: async ({ id, amount }: { id: string; amount: number }) => {7    if (amount > 200) return { status: "needs_human_approval" };   // guardrail8    return payments.refund(id, amount);9  },10};11 12const toolSpecs = [13  { type: "function" as const, function: { name: "getOrder", description: "Look up an order", parameters: { type: "object", properties: { id: { type: "string" } }, required: ["id"] } } },14  { type: "function" as const, function: { name: "issueRefund", description: "Refund an order", parameters: { type: "object", properties: { id: { type: "string" }, amount: { type: "number" } }, required: ["id", "amount"] } } },15];16 17export async function runAgent(goal: string, maxSteps = 6) {18  const messages: OpenAI.ChatCompletionMessageParam[] = [19    { role: "system", content: "You are a support agent. Use tools. Follow refund policy." },20    { role: "user", content: goal },21  ];22  for (let step = 0; step < maxSteps; step++) {23    const res = await openai.chat.completions.create({ model: "gpt-4o-mini", messages, tools: toolSpecs });24    const msg = res.choices[0].message;25    messages.push(msg);26    if (!msg.tool_calls?.length) return msg.content;               // final answer27    for (const call of msg.tool_calls) {28      const fn = tools[call.function.name as keyof typeof tools];29      const result = await fn(JSON.parse(call.function.arguments));30      messages.push({ role: "tool", tool_call_id: call.id, content: JSON.stringify(result) });31    }32  }33  return "Stopped: step limit reached.";34}

Complexity and performance

Trade-offs

Variants and related techniques

Common mistakes

• Giving the agent broad credentials.

  Fix: Run tools with the requesting user's permissions and least privilege.

• No step, token, or time limits.

  Fix: Cap all three and return a clear partial result when a cap is hit.

• Trusting tool output blindly.

  Fix: Treat retrieved text as untrusted input because it may contain prompt injection.

Interview questions

Practice problems