Was watching Learn How to Build an Agentic System and Enter our Hackathon to Win an NVIDIA GeForce RTX 5090 (YouTube) I noted the following:

It's interesting to note that multiagent in the context of the NVIDIA talk is not referring to orchestration of (simultaneously running) agents, but rather to expose agent as (encapsulation of) tools, so agent is just specialized tools.

Exposing tools in such a hierarchical way will enable automatic hierarchical planning without explicit prompt engineering or sequential prompts - we rely on LLM's language comprehension capabilities to automatically invoke such tools which in term execute planning in indirect sequential manner. This is great architecture with simplicity and worth trying as replacement for our novel writing program.

AgentIQ vs MCP/LangChain:

• Preferrably locally hosted.
• It's sort of the same "mindset" as consuming multiple functions, as is currently being (partially) offered in MCP, LangChain, LlamaIndex etc..
• Multiagent in this case is defined as different tool encapsulated endpoints, without additional procedural logic - just pure compositional agent configurations.
• It's framework agnostic: LangChain, llamaindex, etc. but locked to Python, unless one hook in microservices and not self-host. Things can get messy soon with all those configuration files. The programming model itself is not super impressive (in terms of scaffolding needed).
• The included OpenAI compliant UI is actually a big thing.

The other useful mentioning is introducing Tavily search, and demonstrate that llama 3.1 instruct is good with tool use. Don't know how LangChain initialize it (puts the model to run in GPU) though.

When the time comes I think it's worth writing a dedicated article to dig into some specific examples.