The Harness Academy
Three paths · one shared foundation
Understand the machinery behind AI.
From absolute zero to Builder, Solution Architect, or Leader — three destination paths over one shared library, every core idea taught with a demo you can poke.
no signup · no API keys · demos run in your browser
E3The Harness
- 0.4 The vocabulary map
- 1.1 The agent loop
- 1.2 The six ways it dies
- 2.1 Durable execution
- 2.2 Sandboxed tools
- 2.3 Memory & context
E3 · lesson 8 of 15
2.1 Durable execution
Checkpoint every step so a crash resumes exactly where it left off — with no repeated side effects.
- Task received
- Draft the summary
- Send the emailside effect
- Mark complete
- Checkpoint written: step 3 of 4 is now durable.
- Process killed. In-memory pointer lost — but the durable log survived.
- Restarted. Read the durable log: 3/4 steps already done. Resuming at step 4 — no repeats.
- Status
- In progress
- Path
- Builder
- Course
- E3 The Harness
- Lesson
- 8 of 15
01 · Sequence
Start at zero. Leave with judgment.
No prior AI knowledge is assumed, and a built-in bridge course means no prior coding is required either. Everyone walks F1–F3 first; then the paths diverge over one shared content library.
See it, don’t just read it. Every core idea comes with a deterministic browser demo you can step through.
Work at your pace. Checkpoints give instant feedback; capstones use published self-scoring rubrics.
Know when to say no. Judgment is a first-class course, not a footnote.
Showing all 3 paths over one shared Foundation.
- E0 Just Enough Code (optional) (shared with Solution Architect)
- E1 Talking to Models in Code (shared with Solution Architect)
- E2 Grounding — RAG (shared with Solution Architect)
- E3 The Harness (shared with Solution Architect)
- E4 Evals & Reliability
- E5 Shipping & Operating
- Capstone Build something real
- L1 AI Judgment (shared with Leader)
- E0 Just Enough Code (optional) (shared with Builder)
- E1 Talking to Models in Code (shared with Builder)
- E2 Grounding — RAG (shared with Builder)
- E3 The Harness (shared with Builder)
- L2 AI in the Organization (shared with Leader)
- L3 Risk, Safety & Governance (shared with Leader)
- S1 Discovery & Opportunity Qualification
- S2 Solution Architecture for AI Systems
- S3 Scoping, Estimation & the ROI Case
- S4 Delivery, Adoption & Living with the Client
- S5 The Vertical Playbook: Contact Center & CX
- Capstone The Engagement Brief
- L1 AI Judgment (shared with Solution Architect)
- L2 AI in the Organization (shared with Solution Architect)
- L3 Risk, Safety & Governance (shared with Solution Architect)
- L4 Leading AI Initiatives
- Capstone The initiative brief
02 · Choose
Choose your direction.
All three paths share the same Foundation. You pick a direction at the gate after F3, but every course remains open and you can switch at any time.
- Builder
Build, evaluate, and operate real AI systems — from “what is AI?” to a capstone you design and defend. The original path, unchanged.
- Solution Architect
Scope, architect, and deliver AI solutions for real organizations. You’ll build first (E0–E3) — you can’t scope what you can’t build.
- Leader
Decide, govern, and lead AI initiatives — no code required. Deciding and governing well is the scarcer skill.
a path is a direction, not a lock · every course stays open · switch anytime
Signature instrument · F1 / lesson 02
Watch a rule emerge.
class A · class B · ringed — still wrong
Nobody tells it where the line goes — the rule emerges from examples. The three numbers it's adjusting? Lesson 2 →
epoch 0 · 9/20 mistakes
03 · Prove
Every lesson is something you do.
- 01Interact. Real models and deterministic systems run in the page, not in videos.
- 02Check. Instant feedback shows whether the idea held before you move on.
- 03Build. A portfolio-grade capstone is scored against a rubric you can see from day one.
A colleague says "our spam filter uses AI." Which family does it most likely belong to?
- Generative AI — like ChatGPT, since it reads the email textYour pick — not quiteNothing is being generated — a spam filter classifies incoming mail; it doesn't produce text or images.
- Classic machine learning — trained on labeled examplesRight answerLearning the spam/not-spam pattern from thousands of labeled examples is textbook classic machine learning.
- Rules-based AI — someone typed in all the spam phrases by hand
- It cannot be AI without a language model reading the messages
The Engineer capstone rubric
Worth building — the F3 write-up holds up to a skeptic
Evidence one page: task-fit, dials, the assist/automate call
Sound construction — structured outputs, retry armor, right-sized models
Evidence code review + the model map
Grounded honestly (if factual) — citations, threshold, refusal
Evidence the Slack test, passed
Evaluated — 15+ golden cases incl. edges & attacks; green baseline
Evidence
npm run evaloutputA caught regression — the suite has stopped at least one bad change
Evidence the red run, kept as a trophy
Operable — traces answer “what happened, what did it cost?” in minutes
Evidence the 2:14am drill, performed
Survivable — degraded mode + kill switch, demonstrated
Evidence break it on camera
Honest write-up — what works, what doesn’t, what you’d do next
Evidence one page, no marketing voice
04 · Before you start
Questions, answered plainly.
05 · Begin
Understanding beats mystique.
Every impressive AI behavior has ordinary machinery underneath. Learn the machinery and you can build with it, improve it, or lead teams that do — without being fooled by anyone’s demo.