Agentic AI

Introduction

Discipline in development process
Evaluate and Error analysis
Can use Agentic AI to build Agentic AI to build a website teach people how to build Agentic AI.
Parallelism and modularization
Text asset only is more reliable
Multi-modal like sound, vision is less reliable but can be more beneficial. (ChatGPT Agent Mode)
Break down human workflow into smaller tasks.
End-to-end testing and component testing.
Use LLM to judge.
Design patterns: reflection, tool use, planning, multi-agent collaboration.
Reflection: ask LLM to evaluate its own work.
Multi-agentic collaboration: different agents with different roles.

Reflection Design Pattern

Reflection with external feedback is more reliable.
Objective evals: code-based, dataset
Subjective evals: LLM as a judge, rubric-based
Run in Docker sandbox

Tools

Functions
MCP

Practical tips

End to end testing
Component testing
Use spreadsheet to trace and do error analysis.
Play with model often
Use different models
Read prompt regularly
Measure cost and latency
High quality first, optimize later

Patterns for Highly Autonomous Agents

Create a plan first then execute
Applied in coding agents
We don't know the process ahead until runtime
Plan step output to json or xml
Plan step output can be code for execution
Multi-agent can be linear or to have a manager agent
Pipeline can be a python script
Deeper hierarchy vs all-to-all communication

Resources

https://learn.deeplearning.ai/courses/agentic-ai (opens in a new tab)

Udemy Multi Agent