The Tool Use Pattern: How AI Agents Actually Work

Agents Are Just Loops

Strip away the hype and an AI agent is a simple pattern: a language model that can call functions. The model doesn't execute code. It doesn't access databases. It outputs a structured request — "call this function with these arguments" — and your code does the rest.

Here's the core loop:

That's it. Every AI agent — from simple chatbots to complex autonomous systems — runs some variation of this loop.

Why JSON Schema Matters

The tool definitions you give the model are JSON Schema objects. They describe the function name, parameters, types, and constraints. The model uses these schemas to generate valid function calls.

{
  "name": "search_customers",
  "description": "Search the CRM by name, email, or account ID",
  "parameters": {
    "type": "object",
    "properties": {
      "query": { "type": "string" },
      "field": {
        "type": "string",
        "enum": ["name", "email", "account_id"]
      },
      "limit": { "type": "integer", "default": 10 }
    },
    "required": ["query", "field"]
  }
}

The quality of your schema directly determines the quality of the agent. Vague descriptions produce wrong tool calls. Missing constraints produce invalid arguments. Your tool schema is the interface contract between the model and your system.

Schema Design Principles

The Execution Layer

The model outputs tool calls as structured JSON. Your code is responsible for:

async function executeTool(call: ToolCall): Promise<ToolResult> {
  const handler = toolHandlers[call.name];
  if (!handler) return { error: "Unknown tool" };

  try {
    const result = await handler(call.arguments);
    return { success: true, data: result };
  } catch (err) {
    return { error: err.message, retryable: true };
  }
}

Multi-Step Reasoning

The real power emerges when the model chains tool calls. Ask "find our biggest customer and check their support tickets" and the model will:

Each step is deterministic code execution. The model only decides which tool to call and what arguments to pass. Your code controls everything else.

The Enterprise Takeaway

Building AI agents isn't about prompt engineering tricks. It's about designing clean tool interfaces, writing robust execution layers, and controlling the loop. The model is a reasoning engine. Your tools are the hands.

Get the tools right and the agent works. Get them wrong and no amount of prompt tuning will save you.