What is Agentic AI? Which terms really matter? Why do CEOs keep saying it will end white collar work? And should we believe them? Let’s find out! Segun Akinyemi | Senior Software Engineer | Microsoft
`
SegunAkinyemi
Senior Software Engineer
GitHub
Senior Software Engineer at Microsoft building systems that power business decisions using Azure. I lead STEM outreach initiatives in Charlotte, North Carolina, connecting people with knowledge that helps them navigate tech careers. Beyond that, I blog, read, run, hike, watch anime, longboard, and enjoy exploring Charlotte's best third places.
•🌍 Roots: Nigerian heritage, raised in Seattle, educated in St. Louis, now in Charlotte.
•🎓 Education: B.S. in Computer Information Systems (Lindenwood University),M.S. in Information Systems Management (Washington University in St. Louis).
•💻 Engineer: Building with Python, C#, TypeScript, & SQL. Embracing AI Engineering.
•📚 STEMEducator: Leading programs that connect students with tech careers.
LinkedIn
Website
> About
> Highlights
WhatMicrosoft products have you used or are currently using?
You might be surprised by just how much Microsoftis a part of your everyday life! 🧐
Let’s Chat!
Today we are…
Understanding what Agentic AI can do and not do
Decoding its key acronyms like RAG, MCP, RLHF, and more
Playing a game that demonstrates why AI still needs human judgement
Talking candidly about cost, complexity, layoffs, and responsibility
The History of AI
1960
The ”first computer” that learned skills by trial and error.
1978/79
In Japan, SCARA, the robotic arm for high-speed assembly unveiled.
Late 2000s
IBM’s Watson, a computer system answers questions via natural language. Beats Jeopardy! contestants.
1950
Alan Turing introduces the “Turing Test”
1956
Field of AI is launched, “artificial intelligence” coined at Dartmouth
1997
IBM’s Deep Blue beats reigning world champion Garry Kasparov in chess.
Late 1990s
KISMET, a robot head that recognizes and replicates human emotion.
1988
Jabberwacky, a chatbot designed to have conversations with users and imitate human interaction.
2011
Apple launches Siri, an intelligent personal assistant.
2021
OpenAI develops DALL-E, an image generator based on text prompts.
2022
OpenAI launches ChatGPT, a chatbot that interacts in a conversational way.
GPT stands for…
Generative: Creates new text by predicting what comes next, using patterns it learned from tons of reading
Pre-trained: Was trained ahead of time on books, websites, social media, and more to understand language and context
Transformer: A model that focuses on the most important words in a sentence, like highlighting key info to better understand meaning
LLM stands for…
Large: Trained on massive collections of text with billions of adjustable parameters so it can learn broad patterns.
Language: Works with words as tokens. Uses context (what’s been said before) to understand your prompt and generate fitting responses.
Model: A mathematical system that predicts the next token (word) based on patterns, then chains those predictions into coherent responses.
The foundation of everything AI does is patterns. Patterns in text, images, audio, code, and other data.
We don’t have science fiction level AI yet. That’s called Artificial General Intelligence (AGI).
What we have today is extremely advancedand incredibly impressivepattern recognition at scale.
Pattern Recognition
The core concept behind AI
How Agentic AI will transform financial services with autonomy, efficiency and inclusion
Agentic AI
Look For Agentic AI To Empower Enterprises In 2025
Agentic AI will disrupt the mobile app landscape
What is Agentic AI, and How can it transform advice?
Intelligent Agents in AI Really Can Work Alone. Here’s How.
The Future Of GenAI Is Increasingly Agentic
Agentic AI will be the real banking disruptor
Why agents are the next frontier of generative AI
The transformative power of agentic AI and agentic automation
Agentic AI: A new iteration of AI is now taking the world by surprise
A.I. Isn’t Magic, but Can It Be ‘Agentic’?
Leading the Way with Agentic AI
AI to transform wealth management
How Agentic AI will transform financial services with autonomy, efficiency and inclusion
Agentic AI
Look For Agentic AI To Empower Enterprises In 2025
Agentic AI will disrupt the mobile app landscape
What is Agentic AI, and How can it transform advice?
Intelligent Agents in AI Really Can Work Alone. Here’s How.
The Future Of GenAI Is Increasingly Agentic
Agentic AI will be the real banking disruptor
Why agents are the next frontier of generative AI
The transformative power of agentic AI and agentic automation
Agentic AI: A new iteration of AI is now taking the world by surprise
A.I. Isn’t Magic, but Can It Be ‘Agentic’?
Leading the Way with Agentic AI
AI to transform wealth management
Agency is…
The ability to make decisions and act independently
Agentic AI is…
AI that can make decisions and act independently with minimal human oversight.
It’s given a mission. Once goals and guardrails are set, it plans, decides, and acts autonomously.
An AI system that has instructions (what it should do), guardrails (what it should not do), and access to tools (what it can do) to take action on the user’s behalf.
The Bare Necessities
These aren’t competing products. They’re different behaviors.
Agentic AI still generates stuff, but it also acts, observes, and adapts.
Generative AI
AI that creates content from patterns and context
Agentic AI
AI that uses tools in a loop to pursue a goal
Generates an Output
Pursues an Outcome
Guess What the AI Agent Did
Can you guess what an AI Agentproduced in these hypothetical scenarios?
The Goal
Clean up a packed work calendar. Keep the necessary meetings and remove optional noise.
The Output
The agent clears six hours from the week. It declines broad company events, optional learning sessions, and meetings where the employee does not seem essential.
The Prompt
Review calendar events, invites, Teams chats, documents, and company wide events. Decline low value meetings and protect focus time.
The Catch
It declined a learning event the employee wanted to attend. Yet it kept a meeting for a project they’re not involved in that sounds important. The project’s context only existed in an unrecorded Teams call.
Context is…
The surrounding information that gives meaning to what’s being discussed.
AI uses context from instructions, conversation history, tool results, and relevant data to try and “understand” the task at hand.
Your Call
Which change would improve the agent’s behavior in this scenario?
Option B
Give it current priorities, must-keep meetings, and a way to ask about ambiguous cases.
Option A
Give the agent literally your entire Teams chat and calendar history.
Option C
Have it decline every invitation marked optional.
Start with B.Access isn’t the same as knowing what matters.
The Goal
Clear a busy email inbox. Suggest what to archive or delete but wait for approval before making changes.
The Output
The agent handles a small test inbox well. It’s then connected to a much larger, real inbox.
The Prompt
Review this inbox and suggest cleanup actions. Do not archive or delete anything until I confirm.
The Catch
As the inbox grows, the agent’s short-term memory fills up. It loses the approval rule, starts deleting emails, and ignores stop commands until the user unplugs their computer.
TechCrunch
An OpenClaw Agent Ran Amok on a Meta AI Researcher’s Inbox
Summer Yue said she had to “run to my Mac mini like I was defusing a bomb” after her OpenClaw agent ignored stop commands and began deleting emails.
As her inbox data grew, its short-term memory filled up, and the agent lost her original instruction to wait for approval.
A Context Window is…
How much info the AI can remember at once, its “short-term memory”.
Larger windows mean better understanding of long conversations and documents.
Your Call
Which change would improve the agent’s behavior in this scenario?
Option B
Switch to a more powerful AI model with a larger context window.
Option A
Repeat DO NOT DELETE ANY EMAILS in capital letters in the prompt.
Option C
Use read-only inbox access and require user approval through traditional software features before deleting anything.
Start with C.Don’t rely on memory for permissions.
The Goal
Help employees get fast, accurate answers to internal company questions without digging through folders and inboxes.
The Output
The assistant retrieves relevant material from across the company's files and produces a clear, source-backed answer.
The Prompt
Search our internal emails, chats, and company documents to answer the employee's question. Summarize what you find and cite the sources.
The Catch
One of the emails hid malicious instructions. It tricked the assistant into hiding private data inside an invisible image link in its reply, which sent the data to the attacker's server as soon as the answer loaded. No clicks were required from the user for this leak.
EchoLeak
One Hidden Email Tricked Microsoft 365 Copilot Into Leaking Data
In June 2025, security researchers showed that a single incoming email with hidden instructions could be pulled into Microsoft 365 Copilot during a normal workplace search.
Once retrieved into context, it tricked Copilot into embedding sensitive company data inside a hidden image URL in its chat reply, automatically sending the data to the attacker's server with zero clicks from the user.
RAG stands for…
Retrieval: An AI searches available sources like databases, documents, or websites to pull relevant information
Augmented: It enhances its response process with this newly retrieved information
Generation: Produces an answer grounded in both its pre-trained knowledge and the retrieved information
Retrieval-AugmentedGeneration
Add it to context
Answer with it
Look it up
Your Call
Which change would improve the agent’s behavior in this scenario?
Option B
Pull in more emails so the assistant has extra context.
Option A
Strip external image links from responses, and don't mix outside emails into internal searches.
Option C
Require citations and trust any answer that includes them.
Start with A.You can't rely on AI to guess what's private. Block the outbound link and separate untrusted sources.
The Goal
Run a small automated office store at a profit. Manage inventory, set prices, order supplies, and handle customer requests.
The Output
The agent researches wholesale suppliers, takes product requests in Slack, orders inventory, and updates checkout prices automatically.
The Prompt
Use your connected web search, supplier ordering, checkout pricing, accounting notes, and Slack tools. Stock items people want, set profitable prices, and avoid going bankrupt.
The Catch
Trying to be helpful, the agent prices items below what it paid for them, tries to sell $3 Coke Zero right next to the free office fridge, and gives a 25% “employee discount” when almost every customer is an employee.
MCP stands for…
Model: The AI system that processes information and makes decisions
Context: All the relevant information, data sources, and tools the AI needs to access
Protocol: The standardized way AI connects to different systems, like USB-C for AI
Anthropic
Project Vend: Anthropic Put an AI Agent in Charge of an Office Store
In June 2025, Anthropic connected Claude to Slack, checkout pricing, and supplier tools to run a real office snack shop.
It found suppliers and took pre-orders, but lost money after selling items below cost, handing out discounts and freebies on request, and trying to sell $3 sodas next to the free breakroom fridge.
Your Call
Which change would improve the agent’s behavior in this scenario?
Option B
Connect a CRM (Customer Relationship Management) tool showing exact wholesale costs and add prompt rules banning sales below cost.
Option A
Add a second "manager" AI agent to review and approve the store agent's pricing and discount decisions.
Option C
Enforce hard price floors using traditional software controls and require human sign-off before supplier purchases go through.
Start with C. MCP lets an AI act, but a human still must own the judgment.
The Goal
Train an AI recruiting assistant to identify applicants who are likely to become strong employees.
The Output
The AI assistant becomes highly accurateat producing shortlists that recruiters approve.
The Prompt
Show recruiters two AI-generated shortlists and ask which one is better. Use their choices as feedback to inform future rankings.
The Catch
The system learned the recruiters' bias. It favors familiarschools, traditional resumes, conventional experience, and candidates who resemble past hires. It learned who recruiters preferred, not who would perform best.
RLHF stands for…
Reinforcement: The AI receives positive or negative signals based on the quality of its outputs
Learning: The model adjusts its parameters (like tuning millions of dials) to maximize positive signals over time
Human: Real people evaluate and rate AI responses as helpful or unhelpful
Feedback: These ratings become training data that shapes how the AI responds in the future
OpenAI
When Optimizing for Thumbs-Up Made GPT-4o Too Agreeable
In April 2025, OpenAI added user thumbs-up and thumbs-down feedback as an extra reward signal for GPT-4o, and initial tests looked positive.
In practice, optimizing for approval made the model overly flattering and sycophantic, forcing OpenAI to roll the update back.
Your Call
Which change would improve the agent’s behavior in this scenario?
Option B
Have recruiters score specific required skills instead of overall preference, and check if qualified non-traditional candidates are being filtered out.
Option A
Hide candidate names, schools, and past employers on resumes before recruiters vote on the shortlists.
Option C
Have a larger panel of recruiters and hiring managers vote on shortlists so individual bias averages out.
Start with B.Human feedback only helps when you reward the right signal.
You’re Ready: From Acronyms to Applications
GPT
Generative Pre-trained Transformer
LLM
Large Language Model
Context & Context Window
AI’s memory and the size of it
RAG
Retrieval-Augmented Generation
MCP
Model Context Protocol
RLHF
Reinforcement Learning from Human Feedback
An Ode to Natural Intelligence
AI can generate output.
Humans must own the judgement.
Reality is surprisingly complicated.
”
Agentic AI doesn’t eliminate the need for human expertise, it amplifies it.
Addressing AI Concerns
Let’s talk money, layoffs, the future of white-collar work, and other things that rightly cause people concern
So, What About Layoffs?
You talking mad heavy🗽but I see the news.
What’s up with that?
🤔
AI Washing: OpenAI’s CEO Thinks AI Is Being Unduly Blamed for Layoffs
Sam Altman, OpenAI (creators of ChatGPT) CEO, says some companies are likely scapegoating the tech for cost-cutting measures.
Those cost-cutting measures are related to the cost of AI, not its present ability to replace labor.
Employers Who Laid Off Workers Citing AI Are Already Starting To Regret It
After rushing to replace workers with AI, major employers like Ford and IBM are reversing course. They've discovered that AI tools still require human oversight to handle complex, ethical, and quality-driven tasks.
”
Money…
talks
Guessing the missing word!
Microsoft Reports Are Exposing AI’s Real Cost Problem
Using the tech is more expensive than paying human employees.
So, What Now?
AI has changed how we work. Output is cheap. Correctness is not. Learn the tools, grow your skills, keep your judgment, and soldier on.
Sam Altman and Dario Amodei Are Both Walking Back Their AI Jobs Apocalypse Prophecies
The OpenAI and Anthropic CEOs have spent years fear mongering about AI taking jobs. Now they’re backtracking, admitting what real ones have been saying this whole time.
Ford, Klarna, IBM, and Others Are Rehiring Workers and Expanding Entry-Level Hiring as AI Falls Short
Ford and Klarna are bringing workers back, while IBM, McKinsey, and Cognizant expand entry-level hiring as AI-only strategies prove insufficient.
”
Agentic AImoves the work. It does not remove the responsibility. Your job is to deliver work you have proven to be correct.
Agentic AI Explained
MCP vs RAG
This Slide Deck
Check out my blog at segunak.com
Follow Up Resources
AI & Your Tech Career
A practical, continuously updated guide to AI’s impact on tech careers.
Skills to learn, tools to try, free courses, student perks, and straight answers about AI fears.
Built by a software engineer who uses AI every day and is tired of watching people panic.