The Agentic Loop — How AI Agents Actually Work

Workflows are great — predictable, fast, cheap. But they break the moment input gets messy or steps change.

“Schedule a meeting” → workflow handles it fine. “Find time when my whole team is free next week and avoid anyone’s focus blocks” → workflow gives up.

That’s where the Agentic Loop takes over.

---

The Pattern

Every AI agent runs on this same loop:

1. Send system prompt + tool definitions + user message to the LLM
2. LLM responds with a tool to call — not text
3. YOU execute that tool against real APIs
4. Send the result back to the LLM
5. LLM decides — call another tool or give final answer
6. Repeat until the goal is done

The LLM is the decision maker. Your code is the executor. The messages[] array is the memory.

---

In Code

while (response.stop_reason === "tool_use") {
  const toolCalls = response.content
    .filter(b => b.type === "tool_use");

  for (const tool of toolCalls) {
    const result = await executeTool(tool.name, tool.input);

    messages.push({
      role: "user",
      content: [{
        type:        "tool_result",
        tool_use_id: tool.id,
        content:     JSON.stringify(result)
      }]
    });
  }

  response = await callLLM(messages);
}

---

The Key Insight — Messages Array

Every turn you send the entire conversation history:

Turn 1: [ user_message ]
Turn 2: [ user_message, assistant(tool_use), tool_result ]
Turn 3: [ user_message, assistant, tool_result, assistant, tool_result ]

The LLM has no memory between API calls. The messages[] array IS the agent’s memory.

---

What I Built — Calendar Scheduling Agent

User types: “Schedule a standup tomorrow at 10am”

Turn 1: LLM → check_availability({ date: "2026-07-16", time: "10:00" })
Turn 2: { available: true } → LLM → schedule_event({ title: "Standup" })
Turn 3: { success: true } → LLM → "Your standup is booked! ✅"

Three turns. Two tools. One goal achieved.

---

Workflow vs Agentic Loop

	Workflow	Agentic Loop
Input	Structured, known	Natural language
Steps	Hardcoded by you	LLM decides
Cost	Near zero	Token cost per turn
Best for	Predictable tasks	Complex, variable tasks

Production sweet spot: LLM to UNDERSTAND → Workflow to EXECUTE.

---

Same Pattern, Bigger Scale

• GitHub Copilot Workspace — reads code, writes fix, runs tests
• Intercom Fin — searches docs, checks account, resolves ticket
• Salesforce Agentforce — qualifies lead, drafts proposal, updates CRM

The difference is scale and observability — not the core pattern.

Building in public — one agent at a time.