i
Guion · beat por beat
~135 palabras · EN
0–5sHook
"At 3 a.m., one of our AI agents tried to mark a job done — that wasn't."
5–14sSetup / Star
"Sixteen agents run our company. They write, they test, they ship — while everyone sleeps. And this one was sure it had finished."
14–26sTensión
"It had the code. It had a passing check. It was ready to close the task and move on. At 3 a.m., nobody's watching. That's exactly when a confident agent is most dangerous."
26–33sEnemy naming
"The problem was never that an agent makes mistakes. The problem is an agent that's certain — with no one to ask."
33–44sMomento ajá
"So it tried to close it. (silencio) And a gate stopped it. Not a person. A rule we wrote for ourselves: no agent closes work without QA proof and a human keyword. It had the proof. It didn't have the human."
Dirección: subrayar con silencio, no con voz.
44–52sPOV / Lección
"I don't trust autonomous AI because it's smart. I trust it because I built the wall it can't walk through. Trust isn't given to an agent — it's designed around it."
52–57sCTA · keyword
"Comment GATE and I'll send you the exact rule. What would your AI close at 3 a.m.?"
ii
4 ganchos simultáneos
todos = el mismo mensaje
🗣️ Verbal · ≤5s
"At 3 a.m., one of our AI agents tried to mark a job done — that wasn't."
👁️ Visual · frame 0
Terminal oscura con la línea TASK → CLOSE? y un STOP rojo → corte a Roberto serio.
🔤 Textual · ≤7 palabras
IT ALMOST SHIPPED THE WRONG THING
🔇 Auditivo
Silencio seco justo antes de "And a gate stopped it."
iii
POV explícito
regla dura del canal
Se opone a
Autonomía total sin guardrails · el "ship it and trust the model".
Apoya
Human-in-the-loop + reglas de cierre escritas (Doble Puerta de Cierre).
Cree
La confianza en IA se diseña, no se asume.
Experiencia que valida
El casi-incidente real de las 3am (anonimizado).
Línea de razonamiento: el error no es el enemigo → la certeza sin supervisión sí → la solución no es vigilar más, es construir el muro.
iv
Título
pendiente validación JuntaYT
Elegido · Ask-styleWhat Stops an AI Agent From Shipping the Wrong Thing?53 ch
AlternoAn AI Agent Tried to Close Work It Shouldn't (2026)51 ch
v
Thumbnail texts
coherencia hook ↔ thumb ↔ título
Card · línea 1
IT SAID "DONE."
Pill roja (ancla de contexto)
AI AGENTS
Estilo
Playfair · dorado #ECC24E · estudio oscuro
vi
Por qué este tema · la Junta
business-advisor → Hormozi
"Todo se reduce a stakes. La IA reemplaza el trabajo sin stakes; apóyate en proof, que reduce el riesgo para el viewer."
— Framework: AI & Stakes + Problemas vs. Oportunidades (no la cosa brillante nueva — dobla sobre la veta que ya gana)
La data real del canal canta un patrón: gana el riesgo, no la utilidad. Los picos son "empresa operada por IA" con drama o stakes. Los pisos son lo táctico-sin-stakes. Este short es la secuela natural del de "1 of 16 makes a mistake" (530 views) — pero ahora con proof real.
An office where no one shows up…
3 710
♥ 188
Can 1 AI Lead 15 Other AI Agents?
1 016
♥ 28
What Does an AI Agent Manager Do?
608
♥ 12
What Happens When 1 of 16 Makes a Mistake?
530
♥ 10
Why Your AI Sales Agent Keeps Failing
10
♥ 1
I Handed My Weekly Report to an AI Agent
42
♥ 1
Ganadores · stakes / empresa-operada-por-IA
Pisos · táctico sin stakes
