Agentic AI for Everyone

AI can already answer almost anything. Intelligence existing in the world is not the same as you being able to use it. This track closes that gap from first principles: what an LLM actually is, how it gets its knowledge, how every agent is made of the same few parts, and where the whole field is heading. Start with one topic, free. Go deeper where it matters and skip what you do not need yet, without losing the map. Building your own agents is optional, and the easy part once you understand them.

What this course covers

• Tool Selection: Choosing the Right One Among Many
• Scoped Credentials & Least Privilege
• The Output Is a Score for Every Token
• The Cost of Remembering Everything
• The Autoregressive Loop
• Read vs Write, Safe vs Repeatable
• Retrieval Quality: Finding the Right Past
• What Belongs in Each Store
• Generalization vs Memorization & Held-Out Evaluation
• Retry, Backoff, Jitter & Retry Storms
• Read vs Write: Permissions, Confirmation, Authority
• The Runtime Equation: Four Levers
• Hallucination: Plausible ≠ True
• From One Call to Many: Sequential then Parallel
• Conflict Resolution: When Two Stores Disagree
• Transaction Boundaries: All or Nothing
• Positional Information: Sinusoidal to RoPE
• Post-Training → The Assistant
• Memory Observability: Seeing What It Remembered and Why
• When One Model Choosing Everything Breaks: Router & Planner/Executor
• When a Job Outlives the Request
• Argument Validation: Valid Shape ≠ Valid Action
• Decoding: Greedy → Sampling → Temperature → Top-k/Top-p
• Conflict Resolution as an Operating Policy
• Attention & Q/K/V: Why Token Mixing
• Benchmarks & Evals
• One Tool Was Easy: Why That Was a Trap
• Serverless: Fast Start, Hard Wall
• Instruction Tuning (SFT)
• Checkpoint & Resume: Session State Outlives the Box
• Scale & Scaling Laws
• Every Step and Every Tool Call Costs Something
• Model Sizes
• Reading It Back: Memory Enters the Next Context
• Who Executes? Propose vs Authority
• Per-Tenant Blast Radius & Cost Attribution
• Injecting Memory & Tool Results as Structured Context
• Prompt Injection & Indirect Injection
• Name, Description, Inputs: The Schema
• Finding It Later: Store Now, Retrieve on Demand