The Simple Loop Behind Every Fancy Agent

Okay so every fancy agent framework still runs the same tiny loop. The model decides what to do, you execute a tool if it asks, you feed the result back, and you repeat until it responds normally.

import OpenAI from "openai";

const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

// Tool definition the model can call
const tools = [
  {
    name: "getTime",
    description: "Returns the current time in ISO format",
    parameters: { type: "object", properties: {}, required: [] },
  },
];

// Actual tool implementation
function getTime() {
  return { now: new Date().toISOString() };
}

async function runTool(name: string, args: Record<string, any>) {
  if (name === "getTime") return getTime();
  throw new Error(`Unknown tool: ${name}`);
}

export async function agentLoop(userMessage: string) {
  const messages: OpenAI.Chat.ChatCompletionMessageParam[] = [
    { role: "system", content: "You are a helpful assistant." },
    { role: "user", content: userMessage },
  ];

  while (true) {
    const res = await client.chat.completions.create({
      model: "gpt-4o-mini",
      messages,
      tools,
      tool_choice: "auto",
    });

    const msg = res.choices[0].message;

    if (msg.tool_calls?.length) {
      messages.push(msg);

      for (const call of msg.tool_calls) {
        const result = await runTool(
          call.function.name,
          JSON.parse(call.function.arguments || "{}")
        );

        messages.push({
          role: "tool",
          tool_call_id: call.id,
          content: JSON.stringify(result),
        });
      }

      continue;
    }

    return msg.content;
  }
}

(async () => {
  const answer = await agentLoop("What time is it right now?");
  console.log(answer);
})();

What is happening

1. The model decides the next step. Each iteration either returns a normal reply or a tool call.

2. If it asks for a tool, you run it. You execute the tool and capture the output.

3. You push the tool output back in. That gives the model the data it needs to finish.

4. You loop until it replies normally. That final reply is your answer.

That is the core of agents. You can add more tools, a step limit, or memory later, but this loop is the foundation.

ciao