[Proposal] RAG should be integrated with supervisor agent

[subbaksh]

The problem

• Current the RAG agent is a separate agent to supervisor. This means all the knowledge for the agents - documentation, playbooks, knowledge graph are all delegated to a separate agent.
• The consequence is that the supervisor agent acts without consulting the manuals and playbooks.
• While this can be fixed with system prompts (e.g. ask the RAG agent for advise on all queries), it would just add latency to all interactions.

The proposal

• Make RAG part of the supervisor agent using a "search" tool that supervisor has access to.
• For GraphRAG, we add a separate node in the agent graph, as "kg_query" (knowledge graph query) node, that has access to graph traversal tools.
• All tools are will be optional and can be disabled with flags

Benefits

• The supervisor agent is smart about what playbooks and knowledge that already exists, before delegating it to other agents.
• Less latency for documentation questions.
• Better integration of RAG for deep agents (in the future)

Drawbacks

• The supervisor agent is more complex (compared to a simple system prompt)