# swarm.arnao.ai — Self-Running Narrated Demo

**Voice:** Byron — first person, confident, technical-but-warm. A builder showing off a tool he made, not a salesperson pitching one.
**Format:** auto-playing product demo. ~90 seconds, 11 beats.
**Read as:** `[SECONDS] · [ON SCREEN] · [NARRATION]`
**Doubles as:** the spine of a LinkedIn post on harness development.

---

### Beat 1 — Hook (0:00–0:09)
- **ON SCREEN:** Dark mission-control console boots. One idle agent node pulsing green, everything else quiet. Title fades in: *swarm.arnao.ai*.
- **NARRATION:** "I got tired of one model doing one thing and calling it a day. So I built a place where I can launch a whole swarm of agents — and actually watch them think."

### Beat 2 — Launch a mission (0:09–0:19)
- **ON SCREEN:** A "New Mission" panel. Byron types a goal: *"Find a mean-reversion trading strategy that beats buy-and-hold on 5 years of SPY."* Guardrail toggles snap on: MAX AGENTS 14 · MAX SPEND $2 · MAX LOOPS 7.
- **NARRATION:** "Every mission starts with a goal and a set of guardrails. Max agents, max spend, max loops. I'm giving it real autonomy — but on a leash I set. That's the whole game: freedom with a hard stop."

### Beat 3 — The swarm wakes up (0:19–0:30)
- **ON SCREEN:** The graph explodes outward — one planner node spawns a fan of worker nodes. Edges light up cyan as agents hand off. A live tail log scrolls on the right: *spawned · researching · backtesting.*
- **NARRATION:** "And there it goes. Fourteen agents, spinning up in parallel, splitting the problem between them. The graph on the left is them talking to each other in real time. The logs on the right are them thinking out loud."

### Beat 4 — The thesis (0:30–0:42)
- **ON SCREEN:** Two lanes highlight. Most nodes glow green and cyan (labeled *cheap · local*); one node glows purple, labeled *Fable 5 · strategy*.
- **NARRATION:** "Here's the trick that makes it cheap. The cheap local models do the volume — the grinding, the backtests, the thousand boring tries. And I bring in Fable, the expensive model, only for the strategic calls: what to try next, when to change direction. Muscle is cheap. Judgment is where you spend."

### Beat 5 — A loop converging (0:42–0:55)
- **ON SCREEN:** Cut to the LOOP screen. Iteration cards stack: a draft scoring 6.2 → 7.1 → 8.4 → 9.1. An evaluator critiques each one, an optimizer rewrites it. The convergence chart climbs toward the 9.0 line.
- **NARRATION:** "This is my favorite part. One agent writes a draft, another one grades it against a rubric, a third rewrites the weak spots — and they just loop. Six-point-two, seven-one, eight-four. Watch it climb toward the bar."

### Beat 6 — It stops itself (0:55–1:04)
- **ON SCREEN:** The line crosses 9.0. A green banner fires: *✓ CONVERGED — stopped early · saved 3 iterations / $0.06.*
- **NARRATION:** "And the moment it clears nine, it stops. It doesn't burn three more loops for a tenth of a point. It knows when it's done. That's the honest trade — you spend a little money to buy a lot of quality, and you quit while you're ahead."

### Beat 7 — The seductive result (1:04–1:13)
- **ON SCREEN:** Back to the mission. A gorgeous green result card: *Sector-momentum + volatility filter. In-sample Sharpe 1.62 · CAGR 18.4% · beats buy-and-hold.* A small amber tag sits on it: *IN-SAMPLE ONLY — UNVERIFIED.*
- **NARRATION:** "And here's what fell out the other end — a strategy that crushes buy-and-hold. Sharpe of one-point-six. If I posted this curve, half of you would want in. Which is exactly why I don't stop here."

### Beat 7b — The verifier kills it (1:13–1:24)
- **ON SCREEN:** The same verifier gate that killed bad code patches runs on the strategy. A second curve — out-of-sample — flatlines next to the seductive one. A pink stamp lands: *KILLED — OVERFIT · in-sample 1.62 → OOS 0.1 · PBO 0.87.*
- **NARRATION:** "Because when I run it forward, on data it never saw, it dies. The harness caught what my own excitement wouldn't. The swarm generated four thousand strategies — and the harness refused to lie to me about a single one. *That's* the product. Not the curve. The thing that kills the curve."

### Beat 8 — The cost (1:15–1:23)
- **ON SCREEN:** The ledger tallies up: *14 agents · 6m 02s · $0.91.* The spend bar sits well under the $2 cap.
- **NARRATION:** "Fourteen agents. Six minutes. Ninety-one cents. That's a research team that would've cost me a day of my own time — for less than a coffee."

### Beat 9 — The bigger idea (1:23–1:33)
- **ON SCREEN:** Zoom out to the full console — swarm graph, loop screen, and ledger side by side.
- **NARRATION:** "But the strategy isn't really the point. This is how I prototype harnesses now. Before I wire an agent loop into something that matters, I run it here, watch where it gets stuck, and tune the structure — not the model."

### Beat 10 — Tie to the thesis (1:33–1:42)
- **ON SCREEN:** A single line resolves on screen: *The model is the engine. The harness is the car.*
- **NARRATION:** "Because I keep learning the same lesson: the win almost never comes from a smarter model. It comes from a better harness around it. A cheap model in a good loop beats a great model on its own — every time."

### Beat 11 — Close (1:42–1:50)
- **ON SCREEN:** Console dims to the idle green pulse. Wordmark: *swarm.arnao.ai — watch your agents think.*
- **NARRATION:** "So that's swarm. A place to launch agents, give them a leash, and watch them think. I built it for myself — but I think you're going to want one."

---

## Quotable lines (LinkedIn-post spine)
- "I got tired of one model doing one thing, so I built a place to run whole swarms and watch them think."
- "Muscle is cheap. Judgment is where you spend." *(cheap models do volume, Fable plans)*
- "You spend a little money to buy a lot of quality, and you quit while you're ahead." *(loops)*
- "The swarm generated four thousand strategies — and the harness refused to lie to me about a single one." *(the two-act kill)*
- "The product isn't the curve. It's the thing that kills the curve." *(verification is the moat)*
- "Fourteen agents. Six minutes. Ninety-one cents."
- "The model is the engine. The harness is the car."
- "A cheap model in a good loop beats a great model on its own — every time."

## Numbers used (kept consistent with loop.html)
- Loop scores 6.2 → 7.1 → 8.4 → 9.1, threshold 9.0, converged loop 4, saved 3 / $0.06 — **matches the LOOP screen exactly.**
- Mission scope (separate scale, deliberately): 14 agents · 6m 02s · $0.91 · under the $2 cap. This is the *whole swarm*; the $0.06 is *one loop inside it.*