Natural language drifts, LLMs are not an exception

Only 1 of the 4 signals driving agentic SDLC is unambiguous:

• Intent signal – what people want, expressed in natural language,
• Specification signal – what must be true, usually semi-formal,
• Feedback signal – what is wrong, mostly natural,
• Environment signal – what changed around, usually semi-formal.

Agentic code generation does not remove drift. It makes it directly executable, because interpretation is no longer absorbed by humans during meetings or reviews.

Collaboration potential between engineering and business stakeholders tracks almost exactly inverse to formality: the more a notation is designed to eliminate ambiguity for machines, the less room it leaves for non-technical stakeholders to participate.

Check the notations your teams use today against the formality degree: User Stories, ADR, EARS, RFC, BPMN, Gherkin, Event Storming, C4. Wherever a business-critical decision still lives in natural language, that is where your agent has the most freedom to guess.