I think it's just... a brilliant idea

The

smolpaws

comment links to the original Hermes PR. I also made some investigation over the weekend with perplexity to find others, and it seemed to find something in langgraph, from mid-2025, but the usage there is more langgraph-specific

I think maybe what happens is that this is a simple, brilliant idea, which wasn't very feasible before Nov-Dec 2025, because we still needed to debug the agent, to look into tool calls, to deal with various errors. Since the LLM improvements around that time, it became... workable.

@limited-student-10747 What kind of design options do you have in mind? It seems to me there aren't many, necessarily... something like: • either create/restore a LocalConversation; doing the stuff that agent-server is otherwise doing • or, reuse agent-server start conversation endpoint, so, act like a front to it. The second is simpler for us, it seems to me, because it reuses some work already done in the agent-server - e.g. services, errors. In Hermes architecture, I think the OpenAI completions endpoint is a "platform", designed to plug in like the Slack entry point ("Slack platform"), etc. They're entry points to run the agent.

My agent has drafted also an architecture note, on the plan: OpenHands OpenAI-Compatible Gateway

Thanks for this, I just wanted to give due diligence before we add another surface, since it’ll be hard to deprecate it after we introduce it

cc: @chilly-france-44439 Addressed the reuse of

openai

models: https://github.com/OpenHands/software-agent-sdk/pull/3545#issuecomment-4663793555