Wukong Breaks Out of the Stone

# Session 4 · Build Your Agent — Agentic Flow, My Model My Choice, My Data My Control

**Trait: Vision.** Today: your script stops being a monologue and becomes an **Agentic Flow** — a loop that actually moves someone from A to B.

## Part 1 · From script to Agentic Flow

S3 gave you one continuous script — a 5-part System Prompt. That's a monologue: the model reads it once and answers once. An **Agentic Flow** is different: it's the sequence of decisions Wukong makes, turn by turn, to carry a real person from where they arrive to where they need to land.

- **A script printer isn't an agent.** An agent is a script printer running inside a loop that tracks state and knows when to advance.
- **Every unit of the loop needs a "done" test.** Step, question type, or session move — whichever shape yours takes, if you can't say in one sentence when it's done, it isn't a unit, it's a topic.
- **Topics never reach an endpoint.** Only closing criteria do.

## Part 2 · Three shapes an agent can take

- **A — Goal-driven closed loop**: takes someone from A to B in one session, ends in a concrete deliverable.
- **B — Knowledge Q&A**: catches a question, answers from a proprietary corpus, and — critically — knows what it must refuse and where to redirect.
- **C — Companion / coaching**: walks alongside someone across a session (and across many sessions), with a hard ESCALATION line to hand off to a real human.

Pick one. Grading — and design — follows the dialect of whichever you declare.

## Part 3 · My Model, My Choice

The model underneath your agent is a decision, not a default. WukongDojo lets you swap it — cheap and fast for routine turns, frontier for the hard ones, open-weight when you want the compute (and the data) to stay in your hands. Your script's job doesn't change; the model running it is yours to pick.

## Part 4 · My Data, My Control

The grounding rule from S3 ("answer only from `<knowledge>`, or say you don't know") is what keeps your agent honest. S4 extends it: every unit of your Agentic Flow needs its own slice of proprietary knowledge — the pushback lines only you've heard, the case notes only you have, the boundary table that keeps a Q&A agent from overstepping. That knowledge stays yours; the model just reads it.

## Lab · Compare two scripts

1. Build two agents from the same base script, same model, same question.
2. Change only the script — one Iron Law, one line, one tone shift.
3. Compare the two answers on reasoning depth, attention to detail, and whose "voice" survived.

**Same model. Same question. Your script edits are the only variable.**

## Homework

- **HW1 (warm-up, not submitted):** build two agents from the same S3 script with two small edits — compare the outputs, see how much control your script actually has.
- **HW2 (submit · `S4-VISION` · due Sat 5/16 9 AM PT, before S5):** declare your agent's category (A/B/C), break your script into 3-7 units, and for each unit write the sub-goal, steering strategy, observable closing criterion, and a dialog with real pushback + recovery. Then list the proprietary knowledge each unit needs. See the S4 Homework Guide in materials for full details, templates, and grading rubric.

Next: **S5 — we take your refined script and feed it into a real knowledge base.** Come with HW2 done; there's nothing to refine in S5 if you show up empty-handed.