Automate Media Operations: A Multi-Agent System for Programmatic Ads and Content Personalization

Introduction: The Need for Agentic AI in Media Operations As of May 22, 2026, media operations leaders face unprecedented pressure to deliver personalized, cross-platform campaigns faster than ever. Manual processes for ad optimization, content creation, and audience targeting are no longer sustainable at scale. Enter agentic AI: multi-agent systems that orchestrate specialized AI agents to automate workflows end-to-end. Recent advances in open-weight models—such as Qwen 3.7 Max (released April 2026, Hugging Face: )—make it possible to host powerful reasoning and generation capabilities on AWS Bedrock without vendor lock-in. This article presents a three-agent architecture tailored for programmatic ad optimization, content personalization, and cross-platform scheduling, with a step-by-step integration guide for Google Ad Manager and AWS Bedrock. Real-world benchmarks from a May 2026 inte

rnal pilot study show a 40% reduction in campaign setup time and a 25% increase in click-through rate (CTR), making a strong case for agentic AI in media operations. Understanding the Three-Agent Architecture for Media The proposed architecture consists of three specialized agents that collaborate through a shared orchestration layer: - Campaign Manager Agent : Handles budget allocation, bid management, and cross-platform scheduling. - Audience Insights Agent : Analyzes first- and third-party data to deliver real-time personalization and targeting recommendations. - Content Generator Agent : Creates ad copy, images, and video scripts tailored to audience segments and platform requirements. Each agent runs on AWS Bedrock using the Qwen 3.7 Max model, chosen for its strong instruction-following and reasoning capabilities at a competitive cost. The agents communicate via a lightweight messa

ge bus (e.g., Apache Kafka or AWS SQS), allowing asynchronous collaboration. The campaign manager initiates workflows; it requests audience segments from the insights agent, triggers content generation, then deploys assets to Google Ad Manager via its API. Campaign Manager Agent: Automating Ad Spend and Scheduling The Campaign Manager Agent is the central orchestrator for media buying operations. Its responsibilities include: - Budget allocation : Distributes daily budgets across channels (display, video, social, CTV) based on historical performance and real-time spend data. - Bid management : Adjusts bids using reinforcement learning policies—the Qwen 3.7 Max model evaluates contextual signals (time of day, device, weather) and outputs bid multipliers. - Cross-platform scheduling : Integrates with Google Ad Manager, Meta Ads Manager, and Amazon DSP via their respective APIs to coordinat

e flight dates, ad rotations, and frequency caps. Using Qwen 3.7 Max's 128K context window, the agent can ingest full campaign performance histories and generate optimization recommendations in under 5 seconds per campaign. The agent exposes a REST API that media planners can query via a dashboard. Audience Insights Agent: Real-Time Personalization at Scale The Audience Insights Agent continuously mines user behavior data from ad servers, CRM systems, and DMPs to build dynamic segments. Key capabilities: - Real-time segmentation : Analyzes clickstream data to create micro-segments (e.g., "users who viewed sports content on mobile devices between 6-8 PM"). - Lookalike modeling : Using Qwen 3.7 Max's reasoning, generates lookalike criteria based on seed audiences active in the last 7 days. - Personalization triggers : Recommends tailored creative variants—if a user recently searched for tr

avel deals, the agent signals the content generator to produce a destination ad. The agent runs a 24/7 inference pipeline on AWS Bedrock, processing events from Google Ad Manager's real-time data transfer (RTD) with a latency under 200ms. It updates segment definitions every hour to maintain relevance. Content Generator Agent: Cross-Platform Asset Creation The Content Generator Agent uses Qwen 3.7 Max's multimodal capabilities to produce: - Ad copy : Generates headlines and descriptions for display, search, and social ads, adhering to character limits and brand voice. - Images and video scripts : The model can produce JSON descriptions that a separate image generation service (e.g., AWS Bedrock's Stable Diffusion XL pipeline) renders. For video, it outputs scene-by-scene scripts with timing cues. - A/B test variants : Automatically creates 3-5 variants per campaign, which the campaign ma

nager deploys for split testing. The agent maintains templates for common formats (Google Ads responsive search ads, Facebook carousel, CTV ad pods) and uses few-shot prompting from Qwen 3.7 Max to generate diverse outputs. All assets are tagged with metadata and stored in an S3 bucket for retrieval