There’s reasoning. And then there’s reasoning.
Reasoning in LLMs. Reasoning in AI agents. Are they the same?
I was in a meeting recently where I realized we were all using the word “reasoning” to mean completely different things.
Reasoning in LLMs. Reasoning in AI agents. Are they the same?
When one says reasoning, I realize now that one person could mean the reasoning that we get when doing chain-of-thought (COT) prompting with any LLM. Another person could be referring to reasoning in LLMs with reasoning (or thinking) built-in like OpenAI’s o3-mini model. Note that such reasoning by LLMs can be hallucinated (see this paper https://arxiv.org/abs/2509.07339).
And someone else could be referring to the action traces in an AI agent.
So I’ve put together a simple notebook and video to explain 4 approaches and show how they differ.
→ Any model with chain-of-thought prompting (gpt-4o-mini + “think step by step”)
→ A reasoning model (o3-mini with built-in extended thinking)
→ A ReAct (Reason & Act) agent (gpt-4o-mini + tools)
→ A Code Agent (gpt-4o-mini that writes and runs Python + tools)
Same question. Very different reasoning (or action) traces. Guess which approach led to very confident reasoning steps but a hallucinated answer? 🤔
In this short 2min video, I walk through the code and each approach and explain how they differ.
Three things you’ll take away from this:
1️⃣ Not all reasoning is the same. Chain-of-thought prompting, reasoning models, and agent action traces are very different.
2️⃣ Tools can compensate for cheaper models. You don’t always need to pay for an expensive reasoning model.
3️⃣ For analytical questions, code beats thinking. Verifiable computation is more reliable than mental math, no matter how smart the model.
If you’re building something with AI and trying to decide between prompting strategies, smarter models, or agentic patterns, this might help you see the trade-offs more clearly.
Link to full code in a notebook below. Open it in Google Colab. Just get your own OpenAI API key.
https://drive.google.com/file/d/1QkcIBsVX4aTcpftvqKfsrkS3MMdJhjUz/view
#AIAgents #AITutorial #AIRiskManagement
I’m still learning how best to make such videos, so pardon the rough edges.