Agentic DevOps: Why LLMs Should Not Make Every Decision Alone

A developer at SFEIR has proposed a hybrid architecture for agentic DevOps that limits how much decision-making is delegated to large language models. The open-source project Esus, formerly known as Agent Smith, forms the basis of this approach, which was explored as part of internal research at SFEIR. Rather than relying solely on an LLM, the architecture wraps the model in a deterministic 'kernel' that handles code generation, AST validation, and contract enforcement. This neuro-symbolic design lets the LLM handle complex, ambiguous decisions while deterministic software guarantees system invariants and correctness. The approach also redefines the role of human oversight, positioning 'Human-in-the-loop' mechanisms as a deliberate layer within the pipeline rather than an afterthought.