1. What an AI agent actually is

A chat completion answers once. An agent keeps going until a goal is met. The minimum agent is a loop with four parts: a model, a tool registry, a memory of the conversation so far, and a stop condition.

- Model: decides the next action. - Tools: the only way the agent can affect the world (HTTP calls, SQL, file writes). - Trace: the growing message list, which is the agent's short-term memory. - Stop: the model returns a final answer, or a max-step budget is hit.

ts
type Msg = { role: "system" | "user" | "assistant" | "tool"; content: string; tool_call_id?: string };

export async function runAgent(goal: string, tools: ToolRegistry, maxSteps = 8) {
  const messages: Msg[] = [
    { role: "system", content: "You are a careful research agent. Use tools before answering." },
    { role: "user", content: goal },
  ];

  for (let step = 0; step < maxSteps; step++) {
    const res = await chat(messages, tools.schemas);      // one model call
    const calls = res.tool_calls ?? [];
    messages.push(res.message);

    if (calls.length === 0) return res.message.content;   // final answer

    for (const call of calls) {
      const out = await tools.run(call.name, JSON.parse(call.arguments));
      messages.push({ role: "tool", tool_call_id: call.id, content: JSON.stringify(out) });
    }
  }
  throw new Error("step budget exhausted");
}

Everything else in agent engineering — planning, multi-agent teams, MCP, guardrails — is a refinement of this loop. Always build the loop first and add structure only when a real failure demands it.

Exercise: implement `runAgent` with two tools (`search`, `calculator`) and log every step. Ask a question that needs both. Confirm from the log that the model called the tools rather than guessing.

Read in another language

How to study this topic

Scoring

Check your understanding

Badges