AgentsSDK with ResponsesAPI is powerful, however, all the current prod implementations that we have are of chatCompletionsAPI and at minimum I want to port the Langchain components over and refactor the RAG Chatbot using AgentsSDK (natively use ResponsesAPI to get state management and streaming natively). This is proving to be a tedious effort with the examples so far being limited to simple use cases. Think of this as a major upgrade that most “static” RAG chatbot type applications will have to pursue if they have to become “agentic” chatbots. For most production applications, there are external vector stores and Langchain helps in orchestrating the basic flow as “runnables” natively. What I am noticing is that doing a complete rewrite is an R&D effort { ~2 months} including evals delaying other priorities for enhancing the product and introducing new products. Looking for pointers on a more comprehensive guide or examples.
Is that wise? Chat Completions is an industry standard and you are painting yourself into a proprietary corner?
I am unclear on what you mean by proprietary corner ? Also it’s tangential to my original question. Proprietary is where the new AI Platforms ( LLM OS concept) is headed - OpenAI, Google & Anthropic and we can use any model that fits our usecase/s. We are already “proprietary” if we place our bets on OpenAI and frankly the ease of access to developers is unmatched. More specifically, between the API choices, it’s a decision that needs to happen with any software upgrade - responsesAPI offers a few capabilities natively which involves less overhead if you are a really tiny startup. AgentsSDK does support different models : Using any model via LiteLLM - OpenAI Agents SDK
However, given where things are at in production for a lot of companies, “upgrading” their LLM stack is complex because of the orchestration - we have OpenAI, Langchain/Langsmith and the interoperability is what I need for getting the best of both vendors.
The Responses API is the proprietary corner.
Currently you can switch models easily to other providers using Chat Completions.
(don’t get me wrong, I’m a huge OpenAI fan and user, but I’m also keen on good strategy)
I’m not sure I’d agree with you. Yes, OpenAI is trying to see if that’s possible.
Seriously, I’m a single developer and have written two Chatbots with web search and web crawl with just Chat Completions (and loads of complex behaviour in addition). It’s a non-issue.
I’m actually a contributor to that project 
(albeit one PR 
)
Actually you are on to something and didn’t get you wrong so all good there! I just noticed this : OpenAI compatibility | Gemini API | Google AI for Developers so my plans change yet again 
and this calls for a heads down hackathon with my team while we support ongoing customers.
Thank you for asking the tough questions. I slightly agree to disagree on the ease of use with responsesAPI, I am a lone developer and cobbled everything together with chatCompletions and hence I don’t want to change it unless I have to! However, a lot of naive software developers - fresh grads/seasoned folks new to AI want to start with something simple to grasp the concepts in detail so it’s a human / talent question - there is a huge gap between folks who started developing since gpt3.5 vs now and that slows down momentum in going to market and getting revenue.
For sure it takes longer and Responses API is super convenient 