ID ↗
CS 319 · text: pangmandorin.kamil.web.id

CS 319 · Meeting 10 of 10 · capstone

The AI-Native
Company

Everything from nine meetings, applied to one real organization — and audited like you mean it.

Agenda · 100 min

  1. The Kirana arc: 12 weeks, before → after (15')
  2. Fleet audit: live exercise on their data (15')
  3. Capstone brief + rubric (10')
  4. Team presentations: 7 min + 5 Q each (60')

Case study · part 10 of 10 · the arc

Studio Kirana, before → after

MetricWeek 0Week 12
Time buckets (team avg)61% type · 24% review · 8% decide18% type · 31% review · 38% decide
Loops in production09 enabled · 3 paused/retired
Dependency Fridays (Rasa)90 min manual15 min review-only
Specs written per feature0 (tickets only)6 documents, council-produced
Escalations that changed outcomesn/a1 gate + 1 useful checkpoint / day
The line that mattersThe team didn't shrink. The team's decisions per hour tripled.

Their three paused loops are on the slide on purpose. A fleet that never retires anything isn't being audited.

Live exercise · 15 minutes

Audit Kirana's fleet. Three columns.

LoopRunsLast landedQueue @ gateDashboard
L1 triage2203 days ago2 draftsgreen
L2 changelog14021 days ago45 draftsgreen
L3 refund-docs61yesterday14 waitinggreen
L4 seo966 days agoprs_unmerged=7green
L5 design38mixed7 hand-rejectedgreen
KILL
DEFER
SHIP / KEEP

Assign each loop a verdict. For L2: the CLI bug skipped every multi-line draft and logged ok 140 times — which observability rule (M7) would have caught it in one day instead of three weeks?

The final exam is a poster on the wall

Seven operating rules, whole course

1 Landed, not produced — merged not opened, posted not drafted, paid not shipped

2 Skip ≠ success — different log lines, or dead loops stay green

3 Kill rule — two weeks unlanded → kill / defer / ship. No prompt tuning

4 Gate discipline — queue jam: kill/defer/ship, never another producer

5 Verifier at every handoff — boundaries are API boundaries

6 One shared artifact layer — or the org forgets between turns

7 Every agent has an owner + a sandbox — a role label is not a sandbox

The deliverable · teams of 2 · four weeks

Capstone: the Company Playbook

Pick a real organization — student club, family business, your workplace, or Kirana extended. Produce:

  1. Fleet plan — 5 candidate loops, ranked verifier-first (cheapest honest check wins, not coolest AI)
  2. One full spec — five fields, every criterion mapped to a check (A3 standard)
  3. Three-layer verifier — ordered cheap→expensive, calibrated judge (A5 standard)
  4. Six documents for one feature — one hour each, cross-referenced (A4 standard)
  5. A 7-day mini-run — draft-only, logged: landed column, pause classifications, one snapshot diff
  6. The napkin — nodes, edges, state, the conditional loop-back, human gate + gut-check verdict
  7. Runbook + kill switch — stranger-tested

Presentation · 7 minutes + 5 Q

Capstone rubric · 100 pts

25 Spec rigor — binary criteria, refusal conditions, real boundaries

25 Verifier rigor — layer order, maker-checker separation, calibration

20 Honesty of the audit — dead things named dead, theater named theater

15 Graph judgment — gut check applied honestly (a "keep it a loop" verdict can score full marks)

15 Communication — the one-breath explanation test, live

Bonus 5: a documented kill. Retiring a loop is a rite of passage, not a failure.

Field trip · the industry is building this, now

Case: YC's QM — the harness layer, productized

github.com/yc-software/qm — 14.3k stars, MIT. A "multiplayer agent harness for work": every employee gets a scoped agent (memory, files, permissions, sandbox); teams share rooms; admins set one security posture (strict → auto → dangerous; scopes only tighten).

Map it to this course: crons/watches/webhooks = ch 7 triggers · skills shared by grant = ch 16 SOP graduation with governance · per-scope sandbox = tool allowlists made real · Postgres memory = hot/warm/cold state.

The detail worth a whole lecture: its AGENTS.md orders coding agents to "never self-review in the authoring context — dispatch an independent reviewer that did not watch you write the change." Maker-checker (M5) as merge policy, reviewer node (M9) in the repo's own constitution.

What it doesn't solve: your verifiers, your landed-outcome accounting. The harness is the substrate; this course was the rest.

Where it goes

After this course

The stack keeps climbing

prompt → context → harness → loop → graph. The next rung will also be a naming event. Use the word-versus-shift filter (M9) on it.

Edges between companies

A2A: agents delegating across vendors and orgs — the Linux-Foundation-governed protocol. Cross-system edges are coming to your fleet.

The question that stays

"What is the work you will keep doing when most of the rest is automated?" — ch 18's final question. Answer it with evidence, and revisit yearly.

The course in one breath

Agents write the code; the foreman writes the spec, the verifier, and the loop — and audits what lands.

CS 319 · end

You have reached the end
of the course — for now.

The loop is yours. When the work forces a graph, you will know.

Textbook

pangmandorin.kamil.web.id
EN · ID, 19 chapters, free

Keep sharp

Appendix A (prompt library) · B (reference architectures) · C (glossary)

Your final artifact

Not a grade. A fleet you audited, a napkin you can defend, and a runbook a stranger can follow at 3am.

Decks are forkable static HTML — teach this course yourself. That's rather the point.