Why AI Agents Are Breaking the Well-Architected Framework's Core Assumption

For over a decade, cloud architects have relied on the Well-Architected Framework (WAF) to assess system designs by scrutinizing execution paths, failure modes, and trade-offs across pillars defined by AWS, Google, and Microsoft. The framework worked because engineers could diagram every request's flow in advance, even across hundreds of services. Agentic AI systems fundamentally challenge this assumption: a single request can trigger a variable, sometimes unbounded, number of model calls, tool invocations, and inter-agent messages, with the model deciding each next step at runtime. Persistent memory adds further unpredictability, as an agent's behavior may depend on context retained from prior sessions that architects never observed. The core protocols and individual components remain securable, but the dynamic call graph — constructed during execution rather than before deployment — exposes a gap that traditional WAF reviews were never designed to address.