>>> Overwhelming FlashNews >>>
➢ “Only 5% of AI agents monetize”
➢ “Whoever tells you that OpenAI's 𝐀𝐠𝐞𝐧𝐭 𝐁𝐮𝐢𝐥𝐝𝐞𝐫
released yesterday killed Zapier, n8n, Make and co
has not touched any of these platforms.” Link
🡪
➢ Agentic Design Patterns
➢ 82% accuracy looks great in a benchmark. In real
life, Google warns, it still ends in failure if you chase
demos over discipline. Link
🡪
➢ 𝗧𝗵𝗲 𝗡𝗲𝘅𝘁 𝗖𝘆𝗯𝗲𝗿 𝗕𝗿𝗲𝗮𝗰𝗵 𝗪𝗼𝗻'𝘁 𝗦𝘁𝗮𝗿𝘁 𝘄𝗶𝘁𝗵
.
𝗖𝗼𝗱𝗲 𝗜𝘁 𝗪𝗶𝗹𝗹 𝗦𝘁𝗮𝗿𝘁 𝘄𝗶𝘁𝗵 𝗮 𝗖𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻
4.
Power and themoney, money and the
power
Minute after minute, hour after hour
Everybody's runnin', but half of them
ain't lookin'
It's going on in the kitchen, but I
don't know what's cookin’
Coolio - Gangsta's Paradise (feat. L.V
September 2025
7.
…and I pray,oh my god do I pray,
I pray every single day
For a revolution.
4 Non Blondes - What's Up (Offic
ial Music Video)
>>> Tech Evolution >>>
Pay the tollto the angels
Drawing circles in the clouds
Lady Gaga - Abracadabra (
Official Music Video)
LLM (Large Language Model): The core engine trained on massive text
data to understand and generate human-like language. It powers most
modern AI agents.
AI Agent: A system that uses reasoning, planning, and decision-making
to perform tasks autonomously. It often wraps around an LLM to add
goal-oriented behaviour.
RAG (Retrieval-Augmented Generation): Enhances LLMs by retrieving
relevant external documents at runtime to improve accuracy and
context. Ideal for dynamic or domain-specific queries.
Fine-Tuning: A method to adapt a pre-trained LLM to specific tasks or
domains by training it further on curated datasets. It improves
performance but is resource-intensive.
MCP (Model Context Protocol): A newer framework that allows models
to access structured context (like PDFs or databases) in a standardized
way, improving multi-step reasoning and document interaction.
>>> Definitions >>>
Agentic AI
Ramp-Up
Program
1. Intoand Teasers for Agentic AI Ramp Up
2. Prompting Techniques for Better LLM Control
3. RAG
4. Tools, and MCP for Enhanced Context
5. Building Autonomous AI Agents and Systems
6. Agent to Agent Communication and
Collaboration
7. One Workshop to rule them all
8. Creating Effective Agents within ZI Environment
15.
Multi
Agents
Don't push metoo far, don't
push me too far. Tonight,
tonight.
U2 - The Unforgettable Fire (Offi
cial Music Video)
in the contextof AI, the technique
called Retrieval-Augmented Generation
, which enhances large language
models (LLMs) by retrieving
information from an external
knowledge source to use in generating
a response.
"RAG" can refer to a torn piece of cloth or,
19.
Types of
RAG
Type DescriptionUse Case
Simple RAG
Retrieves once, then generates a response
based on retrieved documents.
FAQ bots, basic
knowledge queries
Multi-hop RAG
Retrieves multiple documents across steps
to answer complex questions.
Research assistants,
legal/medical AI
Iterative RAG
Refines retrieval and generation in cycles to
improve accuracy.
Scientific writing,
summarization
Hybrid RAG
Combines dense and sparse retrieval
methods for better coverage.
Enterprise search,
customer support
Streaming RAG
Continuously retrieves and updates context
during long conversations.
Chatbots, tutoring
systems
Contextual RAG
Uses user history or session context to
guide retrieval.
Personalized
assistants
Chain-of-Thought
RAG
Retrieves reasoning steps or intermediate
facts before generating final output.
Math solvers, logic-
based tasks
Tool-augmented
RAG
Integrates external tools (e.g., calculators,
APIs) during retrieval.
Finance, engineering,
data analysis
Multi-modal RAG
Retrieves across text, images, or audio
sources.
Visual Q&A,
multimedia search
Domain-specific
RAG
Tailored to niche datasets like legal,
medical, or technical corpora.
Expert systems
Combines retrieval with autonomous agent Complex workflows,
Who can tellthe difference between…
RAG (Retrieval-Augmented
Generation) Chatbot
● Core Function: Combines a large language
model (LLM) with an external knowledge
base (documents, web) to provide accurate,
source-cited answers.
● Process: Retrieves relevant data first, then
generates a response based only on that
retrieved context.
● Capability: Great for Q&A, internal
knowledge bases, fact retrieval (e.g.,
"What's our PTO policy?").
● Limitation: Cannot perform actions or plan
beyond retrieving and generating text.
Tool-Augmented Chatbot
● Core Function: Extends RAG by
connecting the LLM to external tools, APIs,
databases, or code execution.
● Process: Can use retrieved info and trigger
external functions (e.g., search the web, run
code, fetch live data) within a conversation.
● Capability: Performs specific actions like
booking appointments, checking real-time
stock prices, or running calculations.
● Limitation: Lacks complex, autonomous
planning or self-correction; typically follows
direct tool-use instructions.
AI Agent
● Core Function: A higher-level system that
reasons, plans, and acts autonomously to
achieve complex goals.
● Process: Decomposes goals, selects and
sequences tools (including RAG & other
tools), executes steps, evaluates results,
and self-corrects (learning loop).
● Capability: Orchestrates multi-step
workflows, makes decisions, and learns
from interactions (e.g., "Plan a business trip
and book flights/hotels").
● Key Difference: Proactive, goal-oriented,
and autonomous, unlike the more reactive
nature of RAG or tool-augmented systems.
Choose RAG whenyour primary goal is to provide up-to-date, factual
information to the LLM from a specific, dynamic data source like a company's
knowledge base.
Select Fine-tuning when you need to change the model's behavior, style, or
reasoning to master a specialized domain like coding or legal terminology
Prompt Engineering Vs RAG Vs Finetuning Explained Easily
29.
Context Engineering
Context engineeringis the strategic design of data,
workflows, and environments to provide AI models
(especially LLMs) with the optimal, relevant information
needed for accurate, reliable, and useful responses, going
beyond simple prompting to manage memory, external
knowledge, and user specifics. It involves curating
conversation history, user profiles, documents, and real-
time data to guide the AI's understanding and behavior,
ensuring enterprise alignment and scalability.
30.
Prompt Engineering
Prompt engineeringis the process of designing and refining
inputs (prompts) to guide generative AI models, such as
Large Language Models (LLMs), to produce accurate,
relevant, and high-quality responses. It is a multidisciplinary
field that combines linguistic skill, logic, and experimentation
to "program" AI using natural language instead of traditional
code.
We'll always be together,
Together in electric dreams…
The Leaf People in lockdown do T
ogether In Electric Dreams Giorgi
o Moroder & Phil Oakey cover
31.
LangChain Key Features
•LangChain manages applications' interaction
with different LLMs.
• Chains sequence actions to achieve specific
complex goals.
• Retrieval (RAG) connects LLMs to external data
sources for context.
• Agents are reasoning engines that select and
use various tools.
• Memory allows applications to remember past
conversation context.
32.
LangGraph Key Features
•Models applications as a state machine
for complex workflows
• Uses Nodes to represent specific tasks
and functions
• Edges define conditional paths and
dynamic decision-making
• The State object stores shared memory
and centralizes information
33.
LangChain LangGraph
LangChain's featuresare in several
modules:
● Model I/O: Manages how
applications interact with
different LLMs.
● Chains: Sequences of actions
that connect multiple parts to
achieve a specific goal. An
example is summarizing a
document and then translating
it.
● Retrieval (RAG): Tools that
connect LLMs to external data
sources. (e.g. PDFs, databases,
and web pages) to provide
context-aware responses.
● Agents: Reasoning engines
that use an LLM to decide
which tools to use. (e.g. can be
a calculator or web search).
● Memory: Tools that allow
applications to "remember"
previous parts of a
conversation. This enables
coherent multi-turn dialogue.
LangGraph models applications as
a state machine using three
primary building blocks:
● Nodes: Functions that
represent specific tasks, such
as calling an LLM, searching
a database, or performing a
calculation.
● Edges: Define the paths
between nodes. They can be
conditional, allowing the
system to decide the next
step dynamically based on
the current state.
● State: A centralized object
(often a TypedDict) that stores
the shared memory of the
entire workflow. Every node
can read from and update
this state.
36.
Study!
(things are movingfast)
• NVIDIA turned AI learning into a zero-c
ost masterclass. Here are 8 legit course
s with no free trial traps or paywall gim
micks
• Free ZI:
• Google skills
• Linkedin learning
• Newsletters (choose only one)
• https://www.foma.news/
• linkedin - #gen AI
• Youtube
• Langgraph Tutorial
You can sleep forever, but still you
will be tired
You can stay as cold as stone, but
still you won't find peace
Faithless - One Step Too Far (Offici
al Video) ft. Dido
37.
Building the coreof an agent - Technologies
● LangGraph
● LangSmith
● (Vertex AI)
● Agent-to-Agent (A2A)
protocol
● OpenAI, Antropic
and Microsoft agent
builders