Translated by AI:
At present, libraries such as Langchain and Autogen have emerged for rapid agent development, and they are relatively straightforward to use. When utilizing these libraries, it seems we seldom pay attention to prompt techniques like CoT, ToT, and ReAct. Is it truly unnecessary for us to focus on these? Do the experts have any advice for beginners? I am immensely grateful for your generous suggestions!
“Models” from the o1 and o3 series are systems that directly implement the CoT pattern. Deepseek’s R1 variants are another series of model that lend themselves very well to CoT, but don’t force it. Anthropic’s models are also fairly CoT oriented, and have been for a while even before OpenAI did it. GPT-4.5 is also very CoT-friendly, but it requires its own harness.
So if you’re just starting out, you’re probably fine using o1 or o3, and then optimize from there. The only thing you need to be careful of is that raw models (R1, 4.5, etc) may still need additional programmatic guidance.
