Forge — Raise Your Wukong Well

# Session 5 · Forge — Raise Your Wukong Well

**Trait: Refinement.** Today: your script stops being something you *wrote* and becomes
something you *forge* — round by round, from real runs, until the agent actually does
what you want.

## Part 0 · The power, and the beauty

A neurosurgeon at Peking Union, self-taught in math, wrote ~20 lines of instruction,
let Claude run ~16 hours, and the **Crouzeix conjecture** — open for 22 years — was
proved. Even the problem-setter signed off: correct.

- **Power** — ~20 lines of instruction cracked a 22-year-old math problem.
- **Beauty** — the instruction *itself* is beautiful, in structure and in meaning.

The script is not paperwork. A well-forged script is the most powerful — and the most
beautiful — thing you'll write this course. (Want to dig in? Search: *Crouzeix
conjecture* · GitHub: `jinshanmu/CrouzeixConjecture`.)

## Part 1 · 88% die on the way to production

**88% of AI agents die on the way to production** (2026 · replicated across Anaconda,
Forrester, a16z, MIT Sloan CIO panel). Not because the model is weak — because of
three failure modes:

- **85% × 10 steps = 20%.** Each step looks solid; chain ten together and reliability
collapses.
- **Nobody governs context (Context Debt).** The same question is right today, wrong
tomorrow — the agent never remembers the world changed.
- **It won't say no.** It takes everything, takes out-of-scope, never escalates —
and festers until it bursts.

## Part 2 · Three walls → three paradigms

Nobody is a prophet. We hit three walls and iterated out three engineering paradigms:

- **Prompt engineering** — "just write it well." → *Not enough.*
- **Context engineering** — "just feed it all." → *Still not enough.*
- **Harness engineering** — "put a human on the constraints." → *You own the constraints.*

**The Harness is the runtime constraint layer:** `domain knowledge + an interface into
it → reliable behavior in your domain`. The two shapes a constraint takes: *what the
agent knows*, and *how it is allowed to use it*. The industry loop nobody skips:
**1 · Record** (every time it swings the staff wild) → **2 · Constrain** (turn it into a
rule, feed it back) → **3 · Refine** (keep iterating from real scenarios). Domain
grounding is everyone's opportunity.

## Part 3 · Four actors — all four must move

Raising a Wukong takes four, and missing any one means it never grows:

- **LLM** — the base model. Keeps iterating on its own; stronger every generation. We
don't touch this.
- **Platform (WukongDojo)** — memory · harness · Wukong Studio, all forged under
three-party participation.
- **Script (yours)** — your screenplay: sub-goals, moves, done-when, callback loops.
High density is the point.
- **You** — the human in the loop. You don't write it and walk away — you *raise* it.

## Part 4 · Write a better script — the eight slots

A dense script packs eight things into the context window at once — this is how you add
constraints and make the agent *yours*:

**Role** (who you are) · **User** (who the user is) · **Memory** (what happened before) ·
**Knowledge** (your own knowledge) · **Goal** (what the goal is) · **Rules** (how to do
things) · **Constraints** (hard limits) · **Tools** (tools).

## Part 5 · Measure — today and next

- **Today · view the conversations.** Read every turn your agent had with a real user.
Where it drifts, tighten the script or add a constraint.
- **Next release · robot testing.** Automated agents that probe yours, replay real
traces, and score behavior — so measurement runs in the loop and catches regressions
before they ship.

What gets measured gets raised — today by eye, tomorrow by robot.

## Part 6 · One loop rules them all

You no longer sort flows by type. Describe the flow; the model does the planning; the
loop does the work. **Workflow** (closed loop), **Constraint** (hard limits), and
**Knowledge** (your KB) all fold into a single **Agent Loop**:

> **plan → act → observe → repeat**

Each model generation plans better, uses tools better, and self-corrects better — so the
loop, not the flowchart, is the capability that gets things done.

## Lab · Get hands on (run it live, don't watch a demo)

1. **Message channels** — reach your Wukong where your users already are: Discord,
Telegram, Web. Same agent, three doors.
2. **Conversation & memory** — see every turn your Wukong had with each user: what it
remembered, what to feed back.
3. **Token price & premium** — see what each call costs, top up a little, unlock premium
models like Opus / GPT-5.

**Pricing is transparent** — like a phone card, top up and use what you top up (multiples
of 9: $9 / $27 / $54). Charge = actual model cost **D** + WukongDojo platform fee **9% × D**
→ total **D + 9%×D** (e.g. D = $1 → you pay $1.09). Invite fewer than 5 users · free.

## Homework

- **HW (submit · `ENS5-FORGE` · due Sat 8/29 9:00 AM PT, before S6):** pick the ONE
place in your S4 script that most needs changing, actually change it on WukongDojo
(edit the script **or** feed the KB), run it once on a real scenario, and write a
three-part forge log: **what I changed** (before → after, or filename + 3 entries) ·
**how it went** (2-3 dialog lines + honest effect) · **next step** (what to change
before S6 + one question for Bill). Carry your S4 category (A closer / B Q&A / C
companion). **Forge one deep >> graze seven shallow.** See the S5 Homework Guide in
materials for the full template and grading rubric.

An agent is not a document — it is forged.

Next: **S6 — Demo Day · Info Day.** Every student shows their Wukong live; we talk about
what happens after graduation. Come with your forge log done — S6 we walk it toward launch.