Conference Paper — NeurIPS Marketing AI Workshop 2026
"Automated Content Generation Pipelines: A Comparative Study of Multi-Agent vs. Single-Agent Architectures for SMB Marketing"
García-López A., Martínez R., Chen W. · CULTIVA IA Research Lab, Madrid · Submitted June 2026
1
Figure 1 — Multi-Agent Pipeline Architecture
Score: 8.7/10 ✓ Early stop iter 1
Figure 1. Arquitectura del pipeline multi-agente para generación automatizada de contenido de marketing. El Agente Coordinador (LLM central) orquesta cuatro agentes especializados en paralelo (SEO, Copy, Visual, Analytics). El Agregador fusiona las salidas y el Agente QA evalúa la calidad contra el umbral de marca. Las flechas discontinuas rojas indican el loop de retroalimentación activado solo cuando el score cae por debajo del umbral, implementando la misma lógica de smart iteration que la skill diagramas-cientificos-ia-iterativos.
Iteration 1/2
Model: nano-banana-2 via OpenRouter
Prompt: "Multi-agent AI pipeline architecture block diagram, vertical layers..."
Output: figures/pipeline_arch_v1.png
Quality Review (Gemini 3.1 Pro Preview)
Scientific Accuracy: 1.9/2.0 — Correct agent roles and data flow
Clarity & Readability: 1.8/2.0 — Clear hierarchy, good contrast
Label Quality: 1.8/2.0 — All components labeled
Layout & Composition: 1.7/2.0 — Logical L-R flow, balanced
Professional Quality: 1.5/2.0 — Minor: feedback loop could be clearer TOTAL SCORE: 8.7/10 — THRESHOLD MET (8.0 for conference)
✓ EARLY STOP: Quality 8.7 >= threshold 8.0 — No further iterations needed.
Saved: figures/pipeline_arch_final.png | review_log_fig1.json
2
Figure 2 — CONSORT Participant Flow Diagram
Score: 8.2/10 ✓ Early stop iter 1
Figure 2. Diagrama de flujo CONSORT del estudio comparativo (n=240 SMBs). Los participantes fueron asignados aleatoriamente al pipeline Multi-Agente (n=102) o al Single-Agente (n=102), con n=36 excluidos en la fase de elegibilidad. Las tasas de pérdida en seguimiento difirieron entre grupos (3.9% vs. 6.9%), siendo superiores en el grupo Single-Agente, posiblemente debido a la mayor complejidad percibida de configuración. El análisis final incluye n=98 y n=95 respectivamente.
Iteration 1/2
Model: nano-banana-2 via OpenRouter
Prompt: "CONSORT flowchart for RCT comparing multi-agent vs single-agent pipelines, n=240..."
Output: figures/consort_v1.png
Quality Review (Gemini 3.1 Pro Preview)
Scientific Accuracy: 1.9/2.0 — Correct CONSORT phases, numbers add up
Clarity & Readability: 1.7/2.0 — Clear; minor: font slightly small in exclusion box
Label Quality: 1.7/2.0 — All n values present
Layout & Composition: 1.6/2.0 — Good top-to-bottom flow
Professional Quality: 1.3/2.0 — Phase labels need more contrast TOTAL SCORE: 8.2/10 — THRESHOLD MET (8.0 for conference)
✓ EARLY STOP: Quality 8.2 >= threshold 8.0 — No further iterations needed.
Saved: figures/consort_final.png | review_log_fig2.json
3
Figure 3 — AI Signal Transduction Pathway (Conceptual)
Score: 7.6/10 ✓ Early stop iter 1
Figure 3. Ruta de señalización metaforica del pipeline de IA, inspirada en la cascada MAPK. La "senal de cliente" (brief) activa sucesivamente: Intent-RAS (parser NLP), Orchestrator-RAF (coordinador), SpecialistMEK (agentes especializados), y QA-ERK (evaluacion de calidad) hasta llegar al "nucleo" de contenido publicado. Las lineas discontinuas rojas indican inhibicion por el filtro de compliance de marca.
Improved prompt additions:
+ "Increase ellipse label font to 12pt minimum"
+ "Expand nucleus ellipse width to avoid crowding"
+ "Add phosphorylation step indicators at arrows"
+ "Increase contrast between inhibitor and main flow"
Iteration 2 Review:
Scientific Accuracy: 1.9/2.0
Clarity & Readability: 1.8/2.0
Label Quality: 1.7/2.0
Layout & Composition: 1.7/2.0
Professional Quality: 1.5/2.0 TOTAL SCORE: 8.6/10 — THRESHOLD MET
Saved: figures/pathway_v2.png (final) Total API calls saved: 0 (2 iters needed for Fig 3) Figs 1 & 2 saved 2 calls via early stop.