• Our Multi-Agent Architecture for Smarter Advertising Introduction When we kicked this off, we weren’t trying to ship an “AI feature.” We were trying to fix a structural problem in how our ads business actually runs in software. • On the business side, we have multiple ways of buying-Direct, Self‑Serve, Programmatic-all sitting on top of a mostly consolidated backend. • The infrastructure is shared; the behavior isn’t. • Each buying channel has its own workflows, its own decision logic, and its own flavor of “what good looks like.” On the engineering side, that shows up less as “different stacks” and more as “different brains” wired into the same body: One set of services and data powering multiple buying experiences Channel‑specific flows that encode slightly different rules and heuristics Surface‑specific automation (Spotify Ads Manager, Salesforce, Slack, internal tools) solving overlapping problems in slightly different ways A steady stream of “small workflow tweaks” that are all variants of the same planning / optimization problem, but need to be implemented and maintained in multiple places So even though we’ve done the work to consolidate services, we still end up with fragmented behavior at the workflow layer. • The same core decisions-how to allocate budget, how to choose inventory, how to balance reach vs efficiency vs STR-get re‑implemented per channel and per surface. • Over time, they drift.

Article Summaries:

  • Spotify’s advertising team identified that its backend was consolidated, but the workflow logic for Direct, Self‑Serve, and Programmatic buying remained fragmented, leading to duplicated decision‑making and drift. To address this, they introduced a multi‑agent architecture that adds an intent layer, allowing a single programmable decision engine to interpret advertiser goals (e.g., maximize reach in Brazil while protecting video inventory) and orchestrate existing Ads APIs across all channels and surfaces. Rather than hard‑coding separate workflows or building a rigid rules engine, the agentic approach leverages machine‑learning‑driven forecasting and optimization, ensuring consistent, adaptable behavior across Spotify Ads Manager, Salesforce, Slack, and internal tools.

Sources: