Presentation: Multi-Agent Patterns from Spotify’s AI Powered Advertising Platform

Spotify's Ads Manager utilizes a production-grade multi-agent system powered by Google ADK Java to create AI-powered advertisements. Senior Engineer Pratik Rasam explains the platform's architectural patterns, domain ownership models, deterministic guardrails, and tracing-based evaluation strategies. He also shares hard-learned lessons on agent boundary determination, tool schema optimization, and cost management to avoid monolithic agent pitfalls.

During the presentation, Pratik uses an example of creating an ad for QCon AI to illustrate the platform's capabilities. The advertiser desires to target senior engineers across U.S. tech hubs with a call-to-action to "Register Now." The natural language input is transformed into specific intents, such as audience targeting and geo-location, using an LLM gateway that interfaces with Vertex AI and the internal GCP platform.

The extracted intent is then processed by a multi-agent orchestration layer running on Google ADK Java.

The platform consists of various agents, each responsible for a specific task. For the example ad, these include an Ad Script Generation Agent, an Ad Guardrail Agent to ensure policy compliance, and an Audience Recommendation Agent in pilot. The combined output of these agents is two discrete objects: the audience recommendation and the generated creative content. The architecture enforces explicit ownership, responsibilities, monitoring, and prompt management for each agent, with individual packages and owners.