Agentic AI (Deep Learning AI) by Andrew Ng

• Iterative approach/ multiple steps to complete the task result in a better version of the output compared to a one-shot prompt
• Creating the agent: Decomposing your workflow into discrete steps and asking the agent to execute each of the distinct steps
• Evaluations: score based on objective outcomes (e.g., if the response mentions competitors = negative, …)
• Design patters:
  • (1) Reflection/ critic agent
  • (2) multi-agent workflows
  • (3) error analysis
  • (4) planning
• Using LLMs for comparison (1) Position bias (2)
• Giving expamples is the ‘shots’; so one-shot is one example, two shot is two examples
• Reflection with external feedback improves performance

• Tool calling: execulting python can also be a tool
• Model Context Protocol (clients and servers):
• Evaluating the accuracy of a sub-task e.g.,
• When an agent is not performing up to par, look at the traces of the run (i.e., the intermediate outputs); output of a single steps is called a span
• e.g., find some of the wrong outputs and give the right answer and then give it to the LLM

Improve non-LLM performance

• web search, text retrieval for RAG
  • tune hyperparameters:
    • Web search: number of results, date range, replace
    • RAG: chunk size, similarity threshold

Improve LLM performance

• Improve prompts:
  • Defining the role
  • More explicit instruction
  • Few-shot prompting
  • Decompose the task/ add reflection

Cost and latency improvement