From Vibe Coding to Spec-Driven Development: The Engineering Discipline for AI Agents

Over the past year, AI agents have transitioned from technical exploration into engineering implementation. [1] However, traditional software methods like pass/fail unit testing fail for agents because large language models are probabilistic. [1] Even with a temperature of zero, factors like GPU floating-point non-associativity and routing mechanisms introduce output variations. [1] Additionally, implicit model updates can alter agent behavior without any modifications to the underlying code base. [1]

Initial approaches to generative code synthesis resembled "vibe coding," where developers fed errors back to models. [2] Wrapping this workflow in automated agentic loops allows agents to iterate until tests pass, but these agents can wander when applied to large, critical codebases. [2] To keep agents focused, developers are using spec-driven development to co-create Markdown-based requirements, design documents, and task breakdowns before coding. [2] These documents serve as a shared contract that both humans and agents can read and update. [2]

The transition from linear software engineering to an Agent Development Lifecycle (ADLC) indicates that continuous evaluation must replace static testing. The adoption of spec-driven Markdown artifacts directly addresses the non-determinism and wandering loops that currently plague agentic coding when applied to production systems. Unlike traditional CI/CD pipelines that rely on fixed inputs, this approach enforces rigorous boundaries through shared documentation rather than hardcoded logic. What the sources don't address: how much human oversight is required to validate the agent's changes to the shared Markdown specifications as they iterate.