Agentic AI is defined by its autonomy and decision-making capabilities, distinguishing it from simpler applications like zero-shot prompting, RAG, or fixed LLM chains.
Complex tasks are managed using an orchestrator agent that decomposes the problem and delegates sub-tasks to specialized worker agents like a database agent or web search agent.
An agent operates on a Reason-Act-Observe (ReAct) loop and is built from four key components: the LLM, tools for action, an orchestration engine, and a runtime environment.
The orchestration layer manages an agent's cognitive loop using three memory types: working memory for a single turn, short-term memory for a session, and long-term memory across sessions.
Agents use tools to interact with external systems, with the Model Context Protocol (MCP) standardizing this communication, enabling actions like SQL, vector, and graph database queries.
Developers can build agents at various abstraction levels, from low-level APIs to frameworks like LangChain, LangGraph, and the Agent Development Kit (ADK), up to no-code platforms.
A walkthrough of the Python code for Zooka, an AI cardiologist assistant, demonstrates how to define an agent's profile, instructions, and session management using the Agent Development Kit (ADK).
The Zooka agent demonstrates its ability to analyze user-provided symptoms, use vector and graph searches to identify potential diseases, and recommend relevant diagnostic procedures.
After ending a session and starting a new one, the agent demonstrates its long-term memory by recalling previously discussed symptoms and integrating them with new user input.
Jobs that call for the skills explored in this talk.