AI Agentic Framework

AI Agentic framework - I would proudly call this a "great minds think alike" concept. As I was prototyping this concept, at a parallel point of the same time fabric, Prof. Wang, L had publised a paper[1] on this line. AI Agentic framework would consider a Planner node and one or more worker nodes.

The Planner node:

• Orchestrates interactions with user / system
• Breaks down the task in hand into contained individual steps
• Collates results and workder nodes

The Worker node(s):

• Is initiated for each step outlined by planner node
• Always works in a "function-calling" pattern
• Invokes Tools or external API as needed / defined
• Doesn't get too creative (unless needed) and responds to the point

Key aspects:

• The Planner node is expected to

  • hold the context of the entire conversation across mutiple iterations of tasks sets.
  • Define and establish the system prompts which acts as a north star for workder nodes with each iteration

• The Worker node(s) are expected to

  • Maintain short term memory and for communication with other APIs / systems
  • Maintain a long term memeory if there needs to be multiple calls to APIs / systems (i.e. North Star for the overall context)

Reference:

Lang-chain's ReACT Style Architecture - https://github.com/langchain-ai/langgraph/blob/main/examples/rewoo/rewoo.ipynb?ref=blog.langchain.dev

Citation

[1] https://arxiv.org/abs/2305.04091?ref=blog.langchain.dev