Spotify Engineering shares how they built a multi-agent AI system called Ads AI to solve fragmented media planning workflows across their advertising channels. Instead of duplicating decision logic per channel, they decomposed the media planning problem into specialized agents (RouterAgent, GoalResolverAgent, AudienceResolverAgent, BudgetAgent, ScheduleAgent, MediaPlannerAgent) that run in parallel using Google's Agent Development Kit (ADK) and Vertex AI (Gemini 2.5 Pro). The system takes natural language campaign requirements and generates optimized media plans in 5–10 seconds, down from 15–30 minutes of manual work. Key lessons include treating prompt engineering as software engineering, drawing careful agent boundaries, and grounding LLM outputs with real data via function-calling tools.
Table of contents
IntroductionThe ChallengeThe Solution: Approach and ImplementationResults and ImpactLessons Learned and Future WorkConclusionSort: