Tribeca Media: performance and content AI agents for KEYWORKERS
In short
Tribeca Media is building KEYWORKERS, a platform of AI agents for marketing. Its agent read Meta Ads campaigns but did not act on them. We built the Performance Agent: it reviews Meta and Google Ads, proposes pausing, cutting budget or creating campaigns, executes only when a person approves, and generates new creatives when an ad fatigues. We are now building the Content Agent.
The case on video
Narrated in Spanish.
Starting point
KEYWORKERS gives each client a team of AI marketing agents that work on their real ad and social accounts.
The production system read Meta Ads, but it did not touch campaigns, had no Google Ads and generated no creatives, and there was no way to know how much it spent on AI or when it failed. To serve many more clients, the agents had to be able to act, and do it under control.
Result
0
campaign actions without human approval
What we built
- 01
Diagnosis and phased plan
Technical review of the code and the live system, and a fix plan in blocks, prioritised by risk.
- 02
Observability and AI cost
Health checks with alerts, the cost of every call and image recorded, a spending cap per client and campaign audits in batches at half price.
- 03
Performance Agent on Meta and Google Ads
Reads campaigns, spots what stops working and proposes pausing, cutting budget or creating new campaigns, which start paused. Raising budget is never automatic.
- 04
Human approval
No action runs without sign-off from an authorised person. Approvals expire after 24 hours, every action is logged and a budget cut can be undone.
- 05
AI creatives
When a Meta ad fatigues, it generates a new image with GPT Image 2 and leaves it ready for review. Only approved images go to Meta.
- 06
Content Agent (in progress)
Prepares content pieces with copy and image, carousels and short videos, as previews pending approval. It proposes; it does not publish on its own.
Result
- An agent that acts on real Meta and Google Ads campaigns, not one that only reads them.
- No action on a campaign without a person's sign-off.
- AI cost measured per client, with a daily cap.
- Almost 1,000 automated tests on the system.
- Performance: code complete and being validated with a pilot client. Content: in progress.
Services in this case
Related guides (in Spanish)
Agente de IA para Meta Ads y Google Ads: qué puede hacer
Qué puede hacer hoy un agente de IA con tus campañas de Meta Ads y Google Ads: leer, pausar, bajar presupuesto, crear anuncios y creatividades, y qué no debe hacer.
Agentes de IA con aprobación humana: cómo diseñarla bien
Qué debe aprobar una persona cuando un agente de IA toca dinero o clientes: quién aprueba, cuándo caduca y qué no se automatiza nunca. Con un caso real.
Coste de un agente de IA en producción: cómo medirlo y limitarlo
Cómo saber cuánto gasta un agente de IA y ponerle tope: registrar cada llamada, elegir el modelo por tarea, procesar por lotes y cortar antes de gastar.
Qué revisar en un sistema de agentes de IA antes de escalarlo
Diagnóstico técnico de un sistema de agentes de IA: los nueve puntos que conviene revisar antes de pasar de pocos clientes a muchos, y cómo priorizar lo que salga.
Qué es GTM Engineering y qué significa para una pyme española
Qué es GTM Engineering, qué hace un GTM engineer, en qué se diferencia de RevOps y cómo aplicarlo en una pyme española sin contratar a nadie, con un caso real.
Proyectos de IA a medida para empresas: ejemplos, precio y plazos
Qué es un proyecto de IA a medida, cinco ejemplos reales en empresas españolas, cuánto cuesta y tarda cada fase y cómo evitar que se quede en un piloto.

