Allow the Study Assistant to Hook Into a Custom Assistant

GravityMan

Allowing the Study assistant to consult a Custom Assistant, say "Check with @CarpenterSME for further details on this subject to cross-verify information and create a detailed study plan."

This would allow a user to continue using the Study assistant as intended and to continue being developed independently, while allowing them to consult with SME assistants created by Kagi users with different models and prompts.

While the Study assistant has been extremely useful in its methodology for learning, I've found it to be lacking in depth despite describing the material one is trying to learn.

At times, the Study assistant has been factually incorrect, causing me to revert to a custom SME assistant more often. This not only burns more tokens for the user and Kagi but can also disrupt the learning flow.

Luis

We'd like to further develop the Study Assistant. One of the features we're working on is introducing human collaboration to the Assistant, allowing you to interact with the agent before it produces its final response.

Could you share some examples of how you currently use the Study Assistant (or your custom assistant), particularly cases where it falls short?

(sorry for the slow response here)

GravityMan

I'm unsure what the Kagi Study LLM uses for models or its prompting, so that might be helpful to know, unless it's an in-house secret sauce sort of thing. Same thing with the Coder model.

Anyways, my CISSP study workflow contains the following,

CISSP Tutor - Opus 4.7 (reasoning) - Teach me about XYZ concepts that I'm unsure of
CISSP Research - Research (Experimental) - Check the web in detail for updated material on a topic
CISSP Validator - Opus 4.7 (reasoning) - Review my answers and give me a detailed, harsh breakdown
CISSP Exam Simulator - GPT 5.4 Mini - Rapid-fire exam simulator questions. I'll usually know if something is a hallucination or not, and will usually cross-reference with the Research or Validator agent

I can't give specific examples at this time, but there are times when the Study agent will present me with incorrect or impossible scenarios, along with solutions that are entirely incorrect. While the method used to teach me is much better than the agents I'm using, providing it with the proper guardrails and SME references is not an option.

A good example is with Claude Code, where you plan something out, using either predefined custom agents or having Claude define and create them on the fly, and they collaborate to form the final product.