How much of an AI agent is the framework, and how much is just a loop?

When I started using LangGraph I couldn't tell which part was the framework and which part was the model. To get an understanding, i built a shopping assistant without any framework first. It has two tools, one searches a product catalogue and one fetches the details of a single product. The whole agent is this function: ```python def run_agent(user_message, max_steps=6): messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": user_message}, ] for step in range(1, max_steps + 1): response = client.chat.completions.create( model=MODEL, messages=messages, tools=TOOL_SCHEMAS, temperature=0.1, extra_body=EXTRA_BODY, # provider-specific; empty for non-NVIDIA ) message = response.choices[0].message messages.append(message) # remember what the model said # No tool calls -> the model gave its final answer. if not message.tool_calls: return message.content # Otherwise: run each requested tool, feed the result back, loop. for call in message.tool_calls: name = call.function.name args = json.loads(call.function.arguments) result = TOOL_FUNCTIONS[name](**args) messages.append( {"role": "tool", "tool_call_id": call.id, "content": result} ) return "Stopped: reached the step limit without a final answer." ``` The model only asks for a tool and runs nothing. The loop runs the function, appends the result as a message, and calls the model again. When a response comes back without tool calls, that's the answer. I first read max_steps as a safety net. It's really a budget, because every pass through the loop is one API call. A question that needs five tool uses costs six calls. Then I built the same agent with LangGraph's create_agent. The function above becomes three lines. But print the message list after a run and it's the same sequence: user message, the model asks for tools, tool results, final answer. The framework runs the loop for you. It doesn't replace it. What it does add is memory. The raw loop starts with an empty message list on every call, so "Show me your hiking boots" followed by "Which of those is waterproof?" goes nowhere. With a checkpointer the second question works. Add streaming, retries and tracing, and that's real work you'd otherwise build yourself. My take: write the loop once to learn it, use the framework when you ship. If you start with the framework, the loop is magic, and you can't debug magic. I wrote this up as part of a tutorial that runs on a laptop without a GPU, against any OpenAI-compatible endpoint including Ollama. Link to the article in the comments. Did you start with a raw loop or with a framework? What do you wish you'd known earlier?