Somewhat major architectural proposal of making th...
# proj-agent
l
Somewhat major architectural proposal of making the think tool optional and off by default https://github.com/OpenHands/software-agent-sdk/pull/6159 The motivation behind this is I noticed that deepseek flash was going to think to a very frequently and often overthinking when it used it. I ran a small swebench eval and it indeed reduced cost by 10% will not hurting accuracy.
✅ 1
e
the idea is to trust models built in thinking now that they have gotten so good? That makes sense to me
c
For sure, it’s been many LLM eras… I think the ThinkTool is provided to the agent since Sonnet 3.7, ~March 2025 or so 😅
The funny thing for a while has been that OpenAI models occasionally use it with raw reasoning, apparently - I mean, OpenAI doesn’t send raw reasoning to the user, it sends nothing or a “cleaned up” summary; but occasionally GPTs sent it themselves in the ThinkTool ❤️
Re: the PR, approved I left a tiny comment on a field description One other thought: maybe we can use
cross
tests package and its setup for the test for restore - it avoids mocks, in testing restoring a conversation
l
Hah, LLMs do love to say what isn’t true anymore 😅 thanks for checking! And yeah, this felt a bit nostalgic to me as well!