Klarna's AI system handled 2.3 million customer conversations in a single month, cutting resolution time from 11 minutes to under 2 and saving $60 million in costs. The system runs on a multi-agent architecture built with LangGraph. Meanwhile, Gartner predicts over 40% of agentic AI projects will be canceled by 2027 due to escalating costs and unclear business value.

Google DeepMind's December 2025 research revealed the core problem: unstructured multi-agent networks amplify errors up to 17.2 times compared to single-agent baselines. When agents operate without structured topology, each agent's output becomes the next agent's input, causing errors to cascade rather than cancel. The study tested 180 configurations across 5 agent architectures and 3 LLM families.

Three architecture patterns separate successful implementations from failures. The Plan-and-Execute pattern routes planning to one capable model and execution to cheaper models, cutting costs by up to 90%. The Supervisor-Worker pattern uses a central control plane to suppress error amplification. The Swarm pattern enables decentralized handoffs with guardrails. The key insight: agent count matters less than topology. Structured systems below 4 agents see coordination gains; beyond that, overhead consumes benefits.