Spotify's AI-Driven Ad Platform Automates 70% of Creatives

Call center agent wearing headphones working on a laptop in a modern office setting. Photo: MART PRODUCTION/Pexels

Spotify’s advertising platform has shifted to an AI-driven system that generates ads without human copywriting or music production. Since its launch last year, over 70% of ads now use AI tools, with 20,000 creatives produced for 7,000 advertisers.

The platform uses a multi-agent system where each agent handles specific tasks like script generation, policy compliance, and audience targeting. These agents operate within a shared runtime environment, managed by Google’s ADK framework.

How the System Works

When an advertiser describes their ad requirements in natural language, the system extracts intent for audience targeting and creative generation. For example, “I want to run an ad for engineering leaders across the U.S.” is parsed into technology-focused and U.S.-based criteria.

Agents are designed with clear ownership and responsibilities. Each agent is a separate package with defined dependencies, monitoring, and prompt configurations. This modularity ensures accountability and prevents conflicts between agents.

Challenges and Solutions

Initially, the platform used a single agent for all tasks, leading to brittle behavior as instructions changed. Decomposing tasks into multiple agents improved reliability and isolated issues.

One challenge was geo-targeting ambiguity. For example, “target the South” could refer to various regions. Adding country codes as guardrails improved accuracy.

Another issue was token waste when guardrails flagged issues before generation.

Agent Patterns and Tool Design

Spotify uses several agent patterns: sequential, parallel, and generator-critic loops. Sequential agents pass outputs between steps, while parallel agents operate independently, requiring careful context management.

For example, a geo-targeting tool includes descriptions to prevent misinterpretation. Tools are bundled with agents, ensuring ownership and intent alignment.

Error handling is critical. Tools return structured responses when failing, allowing agents to recover gracefully. For instance, an API failure returns a specific error message instead of a generic server error.

Evaluation and Tracing

Reliability is ensured through tracing and evaluation. Every production call is traced, capturing metrics like duration and tool usage. This data is exported for monitoring and replaying to test model behavior.

Live evaluation runs on a subset of traffic to detect issues in real-time without incurring high costs.

Instead, evaluations are run nightly, and anomalies trigger new assertions to improve future performance.

Spotify’s multi-agent system has evolved from a research demo to a robust architecture, enabling efficient, scalable ad creation while addressing early challenges through modular design and rigorous testing.

Spotify’s multi-agent system emphasizes clear ownership and team-based responsibility. Each agent is owned by a specific team, ensuring accountability and modularity. For instance, the audience resolver agent is owned by a team, while another team owns the creative ad script generation agent.

This ownership model extends to tools, where teams own both their agents and the tools they use. This alignment ensures that tools are tailored to agent intents rather than being generic API wrappers. For example, a tool is designed to meet the specific needs of its corresponding agent, even if it uses the same underlying API as other tools.