The bare essense of Agentic AI in 75 lines of code

[niloy]

Below is the code snippet which implements the minimum code required to get an agentic AI with tool call usage that runs locally.

import ollama from "ollama";

type Message = {
    role: "system" | "user" | "assistant" | "tool";
    content: string;
};

const system = `
You are a genius problem solver. Provide clear and concise answers to the user's questions.
`;
const user = `GOAL: How many r in the word strawberry?`;
const tools = [
    {
        type: "function",
        function: {
            name: "run_javascript",
            description:
                "Run a javascript snippet. Wrap code in an IIFE and return the result. Don't use console.log, just return the value.",
            parameters: {
                type: "object",
                properties: {
                    str: { type: "string", description: "the code snippet" },
                },
                required: ["str"],
            },
        },
    },
];

function run_javascript({ str }: { str: string }): string {
    try {
        return eval(str);
    } catch (error) {
        return `Error executing JavaScript: ${error}`;
    }
}

async function solve(depth: number, messages: Message[]): Promise<string> {
    console.log(`Fetching response from model, depth: ${depth}`);
    const response = await ollama.chat({
        model: "qwen3:4b",
        messages,
        tools: tools,
        stream: false,
    });
    console.log(response.done_reason, response.message, response.message.tool_calls);

    if (response.message.tool_calls?.length) {
        console.log("Processing tool calls...");
        const toolResponse: Array<[string, string]> = response.message.tool_calls.map((tool) => {
            if (tool.function.name === "run_javascript") {
                const result = run_javascript(tool.function.arguments as { str: string });
                return [tool.function.name, result] as const;
            } else {
                console.log(`Unknown tool: ${tool.function.name}`);
                return ["", ""];
            }
        });

        const msg = toolResponse.map(([name, result]) => `TOOL RESPONSE (${name}): ${result}`).join("\n");
        const newMessages: Message[] = [...messages, { role: "tool", content: msg }];
        console.log(`Response from tools:\n${msg}`);

        return solve(depth + 1, newMessages);
    } else {
        return response.message.content;
    }
}

const response = await solve(1, [
    { role: "system", content: system },
    { role: "user", content: user },
]);

console.log(response);