Clean Architecture best practice debate: developers warn that AI agents need security layers beyond standard framework defaults

The discussion centers on treating LLMs (AI systems that generate text) as untrusted external dependencies — similar to how engineers treat databases or third-party APIs — rather than as trusted components of an application.

A key architectural risk: even with well-organized code separation (where the 'reasoning' layer is kept apart from actual tool execution), an AI agent can be manipulated into bypassing security boundaries if the Domain logic lacks strict 'Security Interceptor' layers that validate what actions the agent is allowed to take.

For teams building AI agents and orchestration systems (software that coordinates multiple AI components), this means choosing between building custom security validation layers or relying on built-in filters from frameworks like Semantic Kernel or Agent Framework — each trade-off carries different security and engineering costs.