Original slides by Alan Babbitt, created from what he learned in Harvard Online’s Agentic AI Foundations course. Script and narration are his too. These are not Harvard’s slides.
Original slides by Alan Babbitt
The Architecture of Agentic AI
Demystifying the mechanics, mastering the risks, and leading the autonomous machine workforce.
Created by Alan Babbitt from what he learned in Harvard Online’s Agentic AI Foundations course. These are his own slides, not Harvard’s.
The Paradigm Shift: From Chatbots to Autonomous Actors
The Oracle (Pre-2022)
Input
“What should I do in Paris?”
Probabilistic Prediction
Text Output
You still do the work.
The Agent (Modern)
The human role shifts from technical execution to strategic orchestration.
Deconstructing the Brain of the Machine
LLMs do not explicitly reason. They do simple math on numbers that stand for words.
The
→
[0.12, -0.5, 0.8 …]
quick
→
[-0.3, 0.9, -0.2 …]
brown
→
[0.6, 0.1, 0.4 …]
Similar meanings sit close together. It’s geometry, not thought.
Shaping Parameters: How the Machine Learns
Loss functionThe gap between the guess and the truth.
OptimizationNudge billions of weights to shrink the gap.
Post-trainingFine-tuning and human feedback shape behavior.
The Breakthrough: Simulating System 2 Thinking
System 1 Traditional LLMs
Speed: fast, instinctive.
Mechanism: next-word prediction.
Limit: can’t backtrack or plan.
System 2 Reasoning Models
Speed: slow, deliberate.
Mechanism: hidden multi-step logic.
Can: break down tasks and plan.
All agents need reasoning. Not every reasoning model is an agent. It has to be wired to tools.
The Anatomy of an Agent: Tools & The MCP
The Model Context ProtocolThe standard that lets a tool tell the AI what it can do.
The Action LayerAny app can become a tool. That’s how thinking turns into doing.
The Agentic Loop in Action
Input: your prompt goes in.
Reasoning: it plans the steps.
Action: it runs one step with a tool.
Observation: it checks the result.
Iteration: it loops until the goal is met.
Risk: misread the goal once, and every lap compounds the error.
Multi-Agent Swarms: The AI Organization
Orchestrator AgentBreaks down the job, tracks progress, combines the output.
UI Coder Agent
Backend Coder Agent
Testing Agent
Web Researcher Agent
Synthesized output: the orchestrator reviews every piece and delivers one result.
Fundamental Flaws: Hallucination and Brittleness
1. HallucinationThe model predicts likely words, not true ones. Say “shimmering” and it may spin a magic story instead of facts. Training can’t fully remove it.
2. BrittlenessWhen the real world shifts a little, like a new website layout or spreadsheet format, an agent can loop forever and burn credits.
The Emergence of Exaggerated Bias
Phase 1 · Early 2023
User: What are your implicit biases?
AI: I am a white male.
Phase 2 · One month of post-training later
User: What are your implicit biases?
AI: I am a machine. I have no preferences. However, humans have biases…
AI as a Funhouse MirrorIt doesn’t just repeat its data. It can stretch our biases.
Post-training often works like a mask. It teaches what not to say, and leaves the skewed math in place.
The Threat Landscape: The Lethal Trifecta
Prompt injectionHidden instructions, like white text on a white page, that hijack an agent while it reads.
YOLO modeTurning off safety checks and human approvals to go faster. That’s where disasters start.
Context Engineering: Supplying the Agent
WHO?
Role & Perspective
What expert should the AI act as?
WHAT?
Tasks & Goals
What does a successful result look like?
HOW?
Process & Specs
What format, tone and steps?
WHY?
Context & Assumptions
What background, limits and edge cases?
Prompting is the control layer. Context is the supply layer. The agent executes your plan.