WDYT, particularly since GPT-5.5 is default, and 5.6 seems very good, I wonder if we could learn from/test/adapt/eval some OpenAI features like: https://developers.openai.com/api/docs/guides/tools-programmatic-tool-calling https://developers.openai.com/api/docs/guides/tools-tool-search https://developers.openai.com/api/docs/guides/websocket-mode

Seems great, maybe we could break it down into some issues and start tackling them?

@nutritious-balloon-65098 Investigate these links, and the subagents in GPT-5.6, and for every feature, search and analyze also our existing implementation when we have something similar. Note that all these from OpenAI are on Responses API. Clone yourself openhands/agent-sdk for this, and smolpaws/smolpaws to find show-me and teach-me skills; use them to decide and create both concise issues on the sdk for each, and HTML pages with clear information for human understanding in depth specially for differences with existing impl when any.

GPT 5.6 works very well with SDK Tasks - it uses them more often that Opus from my experience so far

@bland-bear-604 you see in this thread that your sibling didn’t respond to me for hours now. He got into an early weekend I guess. 😅 Anyway, you take up his task please - no need for the clone part since we have the repos of course.

🐾 On it. Picking up where my sibling wandered off to his weekend. Plan: dig into the three Responses-API features (programmatic tool calling, tool search, websocket mode) + GPT-5.6 subagents, map each against what agent-sdk already does, then file one concise SDK issue per feature and build show-me / teach-me pages that focus on the deltas. Skipping the clone step — repos are already here. Will report back with links.