Good read. Indeed about an year ago, when OpenAI’s Responses API was still young, I admit I had strong feelings on the ‘encrypted reasoning’ and stateful-by-default anti-features. They’re provider lock-in “features”. Fast-forward to today, these things have only increased; with Anthropic on encrypted signatures and changing LLM’s history without telling the user; with Google’s Interactions API; with the new OpenAI multi-agent API where even the task sent to the sub-agent is not sent to users. 😓 https://earendil.com/posts/session-portability/

Those are all great arguments for using an open harness like OpenHands @limited-student-10747 @prehistoric-father-39041

Yes, though interestingly, I do think it’s about both open harness and open models. It has to be; because something like • the OpenAI encrypted compaction, or • the multi-agent API hiding agent’s prompts, or • the server-side websearch hiding results from the user are solvable by an open harness; but something like • Anthropic inserting prompts when Claude is used via API, which we don’t see; or • OpenAI sending blobs instead of reasoning, so debugging LLM responses is more difficult need open models, simply. 🤔 That stuff is happening behind the LLM API, and the true LLM prompt or LLM response just doesn’t reach us, and it does affect the model’s behavior…