Draft: Initial Agent to Agent system

Replaces current runtime structures with a supervisor-run agent to agent system

NOTE: This also refactors the ontoflow integration to a tool (multigrain).

TODO:

• We're often hitting a GraphRecursionError. While this can be solved trivially by increasing the limit; that's addressing the symptoms, not the problem. I think this can be alleviated by using a planning agent, or moving from a supervisor to a network layout. Or better prompting.
• The Supervisor is way too trigger happy on the RAG, asking it questions it already has the answer to, or things it wouldn't know. This has been addressed on the ontoflow side with the new jupyter_agent. I need to implement these calls as tools. This may also be helped by implementing persistent memory to the Agent(s).
• Models are currently hardcoded. Should be simple enough to change. It would be nice to add a %%lara /model gpt-oss command, for example.

Below is a graph of the current layout:

Where the rectangles are Agents, and the diamonds are Tools