Agents: The Tool-Calling Loop

Decide, act, observe, repeat

An agent is an LLM in a loop with tools. Each turn the model either returns a final answer or requests a tool call; your code runs the tool, appends the result as a message and calls the model again, until it answers or a limit is reached. That is all an agent is; the frameworks supply the loop, tool schemas, message formats and helpers. Use agents when the number and order of steps cannot be known in advance (for example research, debugging, multi-source questions). For fixed sequences, a plain chain is simpler, cheaper and more predictable. Always set a maximum number of iterations, a token and time budget, and log every tool call.

Models that choose actions

An agent loops between thinking and acting; workflows give you control over that loop.

The loop in pseudocode

Frameworks wrap this; knowing the loop helps you debug. Illustrative; not run here.

messages = [system, user_question]
for step in range(MAX_STEPS):                 # always cap the loop
    reply = model.invoke(messages)             # model sees tools via bind_tools
    messages.append(reply)
    if not reply.tool_calls:                   # no tool requested -> final answer
        return reply.content
    for call in reply.tool_calls:
        result = run_tool(call["name"], call["args"])   # YOUR code validates and executes
        messages.append(tool_message(call["id"], result))
return "Stopped: step limit reached"

Quick check: What must every agent loop have?

• A maximum number of iterations and a budget
• An unlimited number of retries
• No logging
• Admin access to everything