They teach you to build an agent.
This teaches you to ship one that survives production.
Evaluation, observability, reliability, security, and governance for real agentic systems — the last mile academic and edtech courses skip. Three modules free.
What you'll actually build
Most courses teach you to build an agent — not to know whether it's actually good, or to catch regressions before users do.
What you build: A production eval harness: golden datasets, an eval-gate in CI, and TDD for agents (Module 3.1).
The artifact: A golden-set pass-rate report that blocks a bad deploy.
Real-world anchor: the instructor's PACCA healthcare platform ships behind a 100-case golden set at a 100% eval-gate pass.
A service fails loudly with a 500. An agent fails silently — it keeps generating tokens, confidently wrong.
What you build: Distributed tracing across LLM calls, tools, and sub-agents; reliability and failure-recovery patterns (Modules 3.2–3.3, 2.4).
The artifact: An observability trace + a multi-agent topology diagram pinpointing where a silent failure propagated.
Covers single-agent, orchestrator-workers, and decentralized handoff topologies most curricula never reach.
Before an agent touches production it has to be secure, governed, and cost-controlled — the part academic courses skip.
What you build: OWASP-LLM-Top-10 threat modeling, guardrails, governance artifacts, and cost/latency optimization (Modules 3.4–3.7).
The artifact: A threat model + guardrail spec and a cost-latency curve for the reference system.
Real-world anchor: the instructor's CRISP-AG framework governs enterprise agentic AI to ISO 42001 / NIST AI RMF.
For each broken prompt, classify its failure mode (from the taxonomy in Segment 1) and sketch a fix. Reveal the answer to self-check, then continue to the next one.
Broken prompt: Tell me about the customer.
Input: a 3-paragraph customer complaint email.
Failure mode: Ambiguous task
State exactly what to extract, with the data delimited: "From the customer message below, extract: name, issue category, and urgency (LOW/MEDIUM/HIGH). <message>...</message>" (Rung 1 + Rung 3).
The full curriculum — 18 modules, ~150 hours
Three modules are free below. The rest — retrieval, agentic orchestration, and production engineering — is the paid depth.
M0: Foundations On-Ramp
M1: Retrieval & Grounding
M2: Agentic Orchestration
M3: Production Engineering
Who's teaching this
David Reed — Head of AI/ML & Agentic Delivery — PhD CS, MBA, Wharton fellow
- 61.4% FAANG placement rate (vs ~5% industry baseline)
- 22% performance lift across 400+ MAANG instructors
- PACCA — production multi-agent healthcare AI, eval-gated, HIPAA / FDA-SaMD
- CRISP-AG — enterprise agentic-AI governance framework (ISO 42001 / NIST AI RMF)
- ALCA & LPA — edtech agents with 20–90% measured time savings
Inventor on the Amazon recommendation-engine patent (US 6,850,988).