The rapid adoption of Large Language Model (LLM)-based coding agents has introduced a paradigm shift in software engineering. However, the current ecosystem heavily relies on source code as the primary artifact, leading to vendor lock-in and systemic vulnerabilities such as context pollution and hallucination in long-horizon sessions. Fundamentally, LLMs are probabilistic generative models; therefore, delegating full control of system state transitions to them compromises deterministic integrity. To address this, we propose the 'Living Functional Design Specification (FDS)' framework, a novel multi-agent orchestration architecture that establishes a dynamic natural language document as the Single Source of Truth (SSOT). By treating source code merely as a reproducible byproduct, this architecture isolates probabilistic chaos. Furthermore, we reject the illusion of full automation by embedding a Human-in-the-Loop (HITL) mechanism at critical state transitions. Human oversight via a deterministic Diff-viewer ensures semantic integrity, while a context pruning algorithm maintains memory efficiency (O(1) active context limit). We demonstrate that an uncontaminated FDS can be reliably extended into heterogeneous assets, such as automated patent claims and academic papers. Ultimately, this vision paper declares that the future of software engineering lies not in writing code, but in defining objective functions and safeguarding the deterministic truth against probabilistic machines.
Jungmin Koo (Thu,) studied this question.