Multi-Agent Retail Media Campaign Optimization: A 10-Retailer Pilot Boosts ROAS by 25%

The Consortium and the Pilot: Why 10 Retail Media Networks Joined Forces As of May 24, 2026, a consortium of 10 retail media networks (RMNs) publicly released the results of what they describe as the first documented multi-agent pilot for campaign optimization. The pilot ran on AWS Bedrock, leveraging specialized AI agents to improve advertising performance across RMNs. The consortium included major retailers from North America and Europe, each contributing anonymized campaign data and operational resources. The goal was to test whether a multi-agent architecture could consistently outperform traditional rule-based or single-model approaches in real-time bidding and targeting. “This pilot was designed to answer a simple question for enterprise operations leaders: Can multi-agent systems deliver measurable ROI in the complex, high-frequency world of retail media?” said a spokesperson for

the consortium. The participants agreed to share results to accelerate industry adoption and create a replicable blueprint. Agent Roles: Audience Targeting with Qwen 3.8 Max and Bid Optimization with Llama 5 The architecture employed two primary agent types, each powered by a different large language model (LLM) selected for its strength in the specific task. Audience Targeting Agent (Qwen 3.8 Max) This agent was responsible for real-time audience segmentation and ad creative alignment. Using Qwen 3.8 Max (released in April 2026 by Alibaba Cloud), the agent analyzed anonymized user behavior data – browsing history, purchase signals, and contextual cues – to predict the most responsive segments for each campaign. Qwen 3.8 Max’s ability to process long-context sequences (up to 128K tokens) allowed it to incorporate recent session data without truncation. According to the , it achieves stat

e-of-the-art performance on multimodal understanding benchmarks, making it ideal for parsing both text and image ad creatives. Bid Optimization Agent (Llama 5) The second agent managed bid price adjustments across programmatic channels. Powered by Meta’s Llama 5 (released in December 2025), this agent processed real-time auction signals, budget constraints, and conversion probability estimates to optimize bids per impression. Llama 5’s advanced reasoning capabilities allowed it to simulate thousands of bid scenarios per second. Per the , its fine-tuning on ad relevance tasks improved decision speed by 40% over previous versions. Both agents ran on AWS Bedrock, which provides managed multi-agent execution environments. AWS Bedrock’s enables synchronous orchestration, allowing the two agents to share a common state and pass intermediate results without latency penalties. Data Pipeline Desi

gn: How Agents Shared Information in Real Time The data pipeline was built on AWS Bedrock’s agent collaboration framework. Key design elements included: - Shared event bus : Real-time campaign events (impressions, clicks, conversions) were streamed to an Amazon EventBridge bus. Both agents subscribed to relevant event types. - Agent context pool : A high-speed in-memory store (Amazon ElastiCache for Redis) held the current state of each campaign, including budget remaining, winning segments, and bid history. Agents read from and wrote to this pool at sub-millisecond intervals. - Orchestrator service : A lightweight orchestrator managed the interaction loop: the audience targeting agent evaluated new inventory opportunities and flagged high-value user segments. The bid optimization agent then received these signals along with real-time auction data to determine bid prices. If a bid was wo

n, the audience targeting agent updated its segment model with conversion feedback. - Fallback logic : In case one agent failed or produced an out-of-range output, a rule-based fallback used historical average performance to maintain stability. This reduced downtime during the pilot by 95%. The entire pipeline was designed for low-latency decision-making (under 200 ms end-to-end), critical for real-time bidding in retail media. Measurable Outcomes: 25% Higher ROAS and 30% Faster Campaign Setup The pilot ran for 10 weeks across 150 campaigns, comparing the multi-agent system against a control group using a single-model approach (a single LLM handling both targeting and bidding). The results, published by the consortium, showed: - Return on ad spend (ROAS) : 25% higher on average for campaigns using the multi-agent setup. - Campaign setup time : Reduced from an average of 3 days to approxi

mately 2 days – a 30% improvement. Automation of audience segmentation and initial bid parameter generation accounted for most of the gain. - Budget utilization : Multi-agent campaigns spent 92% of allocated budget versus 78% in the control group, minimizing underspend. - Conversion rate : Increased