Our Multi-Agent Architecture for Smarter Advertising

Pratik Rasam, Ralph SylvainFeb 2026

Problem

Advertisers on Spotify had to configure every dimension of a campaign by hand: goal, audience, budget allocation, schedule and ad format. That took multiple screens and 20+ form fields. Nothing showed which configuration historically works for similar campaigns, so they guessed, and testing another approach meant starting over. Planning logic was also re-implemented per buying channel and drifted apart.

Example from the article: a media plan took 15–30 minutes to build manually.

Solution in brief

Spotify split media planning into small LLM agents: a router, four agents that turn free text into structured campaign fields, and a planner agent. The four extraction agents run in parallel. The planner turns their output into 1–5 ad set recommendations with a heuristic scoring engine over historical campaign data. Tools return the real data (geo targets, categories, performance), so the LLM decides what to call and does not invent facts.

Questions this case answers
  • How to decompose a multi-input planning workflow into cooperating LLM agents that call existing services as tools and answer within a few seconds
  • How to ground LLM recommendations in historical performance data so that the model chooses actions and tools supply the facts
  • How to turn a multi-step campaign planning form into a conversational flow that returns a media plan in seconds?
  • How to split an LLM workflow into specialized agents so that independent steps run in parallel?

Tools and methods

  • Multi-agent architecture (router + resolvers + planner) — one agent per skill or data source, independent agents run in parallel
  • Function calling — agents get real geo targets, ad categories and historical performance through tools
  • Heuristic optimization rules — score historical campaigns by cost, delivery rate, budget, duration and targeting match
  • Google ADK 0.2.0 — agent orchestration, sessions, tool integration
  • Gemini 2.5 Pro on Vertex AI — language understanding and generation
  • PostgreSQL + in-memory cache — historical performance data
Continue reading — it's free

Google or a one-time email code. No password.