Natural Language Drift in Agentic SDLC: Why LLMs Make Ambiguity Executable

Natural language drifts, LLMs are not an exception

Alex argues that agentic code generation does not remove drift—it makes interpretation directly executable. Only one of four signals driving the agentic SDLC—the intent signal—is unambiguous.

Why This Matters

The collaboration potential between engineering and business stakeholders tracks inversely to formality: the more a notation eliminates ambiguity for machines, the less room for non-technical participation. Wherever a business-critical decision lives in natural language, that is where the agent has the most freedom to guess, turning ambiguity into production risk without human review buffers.

Key Insights

• Intent signal is the only unambiguous driver in agentic SDLC; specification, feedback, and environment signals remain semi-formal or natural (source: dev.to article on natural language drift).
• Formality axis acts as a gating axis: structured enough to be deterministic yet readable enough for operators defines locked-decision schemas in the spectrum between ADR and EARS (source: top comment on dev.to).
• Notations like User Stories, ADR, EARS, RFC, BPMN, Gherkin, Event Storming, and C4 vary by formality; wherever a critical decision stays in natural language is where agents have maximum interpretative freedom (source: dev.to article).

Practical Applications

• {use_case: ‘Engineering teams using Gherkin for acceptance criteria reduce agent guesswork by formalizing behavior’, pitfall: ‘Keeping requirements exclusively in User Stories invites drift as agents misinterpret vague intent over time’}
• {use_case: ‘Architects adopt C4 models to lock structural decisions’, pitfall: ‘Relying solely on event storming diagrams without formal constraints allows agents to infer incorrect boundaries’}
• {use_case: ‘Platforms like Temporal use deterministic workflows to enforce specification signal’, pitfall: ‘Over-relying on full formal specifications excludes business stakeholders from validation loops’}

References: