OpenHands

RFC: the agent-sdk has an `ask oracle` tool, through which the LLM can ask a better LLM for an opinion, a review, or a plan when it thinks it needs it.
This is not surfaced in the UI in any way: the current behavior is hidden behind an `oracle` profile:
• if the user has an LLM profile named "oracle", then the tool is included in the agent's tools, so it can use it
• if it doesn't, nothing tells them about the feature.
My agent proposes a way to surface it: <https://github.com/OpenHands/OpenHands/issues/17217|a greyed-out "oracle" profile>
WDYT?

Wow i didn't know oracle was a thing, seems like this would be nice to have

Do llms actually ask for help though or are they selfish?

It depends on llm and prompt, but if the user knows about this, they can also just say - hmm are you sure, ask the oracle. Or an Agents.md can set the rule e.g. : for difficult tasks ask the oracle.

 :sweat_smile: It depends heavily, this reminds me of Gemini, who doesn’t like other LLMs, one convo when it heard that Claude gave it a bad review, was like “_oh so we are seeing other people now! let me see what Claude pretended to have found… I’m not going to lose this user to Claude!”_ :joy: