CS 319 · Meeting 1 of 10

The Shift
+ Quick Win

The 10x coder is dead. Long live the Agent's Foreman.

Agenda · 100 min

Hook: three years, three roles (10')
Concepts: decisions/hour, three jobs, the loop primitive (35')
Example: a 20-line loop, read line by line (20')
Case: meet Studio Kirana (10')
Quick win: you are the for loop (25')

Reading: chapters 1–3 · pangmandorin.kamil.web.id

Where we are

Three years, three roles

2024

Operator

You type prompts to one model. The skill is asking well.

2025

Manager

You run several agents side by side. The skill is delegation.

2026

System designer

You write the loops that prompt the agents. The skill is the cycle.

"I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops."

— Boris Cherny, Head of Claude Code, Anthropic (2026)

And mid-2026 added a fourth rung: the org designer — wiring loops into graphs (Meeting 9).

Core concept · ch 1

Lines per hour → decisions per hour

Dead metric

Lines per hour. The agent writes a thousand lines a minute. An AI-augmented engineer types fewer lines than the 2015 median.

Live metric

Decisions per hour: which thousand lines should exist, where, and why — expressed as specs, prompts, and verdicts on agent output.

The fourth job that disappeared

Typing the obvious code — boilerplate, CRUD, migrations, the mechanical eight hours. Agents do it faster, in parallel, overnight. The engineer who clings to it loses to the one who lets agents type while they read the plan.

Core concept · ch 1

The three jobs of the foreman

A construction foreman never swings a hammer. Their value is coordination and judgment.

1 · Read the plan

Understand what the system should do — the spec: PRD, architecture, API contract, schema. Old artifact: Jira ticket.

2 · Sequence the trades

Which agent does what, in what order, with what handoff. Old artifact: sprint board.

3 · Inspect the work

Review artifacts, leave coaching notes, graduate notes to playbooks. Old artifact: PR review.

None of the three goes away when AI writes the code. Errors in the spec are now amplified by an agent that happily produces a thousand confidently-wrong lines.

Core concept · ch 3

The loop: the primitive of everything

discover→ plan→ execute→ verify↺

Loop engineering = designing this cycle instead of writing individual prompts. Everything in this course is a refinement of this primitive.

Every loop has 3 moving parts

trigger — cron, webhook, or agent-to-agent: what starts a pass

state — what the loop knows: queue, draft, verdicts

exit — goal satisfied · max iterations · stagnation breaker · human escalation

Example · ch 3, runnable

The smallest honest loop

while true; do
  TASK=$(next_task queue.txt)               # TRIGGER: item appears
  [ -z "$TASK" ] && sleep 60 && continue

  PLAN=$(agent "plan: $TASK" < spec.md)      # discover + plan
  OUT=$(agent  "do:   $PLAN" < tools.yaml)   # execute

  if verify "$OUT" checks.txt; then          # VERIFY: deterministic first
    land "$OUT"; mark_done "$TASK"           # landed, not just produced
  else
    RETRY=$((RETRY+1))
    [ $RETRY -ge 3 ] && escalate "$TASK"     # EXIT: human escalation
  fi
done

Find the trigger. Find the state. Find all four exit conditions. If you can't name them, it isn't a loop — it's a script with hope.

Case study · part 1 of 10

Meet Studio Kirana

Eight people in Jakarta — 5 developers, 1 PM, 1 designer, 1 QA — building CekFaktur, an e-invoicing SaaS with 4,000 SME customers. They "use AI": autocomplete in the editor, plus copy-paste chats at midnight. No specs. Merges at 23:00. QA by vibes.

Their week, measured

61% typing · 24% reviewing · 8% deciding · 7% meetings

Symptom

"We use AI" = autocomplete + pasting. Which era are they in?

This course

We follow Kirana for 12 weeks as they become loop-native. Audit at Meeting 10.

Quick win · 25 minutes · in class

You are the for loop

3' — Pick one small task you did ≥3 times last week (summarize a doc, triage mail, draft release notes).
5' — Write three lines: GOAL, BAR (3 pass/fail checks), EXIT (max 2 tries).
10' — Run any chat LLM as the executor. You are the trigger and the verifier. Score against your bar.
5' — Log: how many decisions did you make? How many lines did you type?

That decisions-to-lines ratio is the whole course.

Assignment A1 · due before Meeting 2

The Day Audit

Task
For 2 working days, mark every 30-minute block: Type / Review / Decide / Other (ch 1's buckets). Then take your largest Type task and write a one-paragraph spec — goal, boundary, 3 pass/fail criteria — and hand it to an agent. Compare outputs.

Deliverables
Audit table (2 days × buckets, %) · the spec · a 200-word comparison: where did the agent succeed, fail, and what was missing from the spec?

Rubric
Honest audit 40% · spec quality (are the criteria really pass/fail?) 40% · insight of the comparison 20%.

Rule of thumb for locating yourself: Type > 30% → operator era. Decide > 20% → system-designer era.

Next meeting

The Council and the Context

Read (≈45 min)

Ch 4 — The Council Pattern: why one brain isn't enough.

Ch 5 — Context Engineering: the art of the right information.

Bring

Your A1 audit, and one real decision you'd normally make alone. You will give it to a council of four agents who disagree with each other.

Course text (free, EN/ID): pangmandorin.kamil.web.id

The Shift+ Quick Win