Behavioral Variability

The emergence of different execution trajectories from the same input to an LLM-backed agent system (sources/2025-fournier-agentic-ai-process-observability).

Following Milani, Dumas, Matulevicius 2012 process-variant terminology, a branching point in a mined agent process model is either a:

Decision point — a variation point whose branches correspond to distinct, specified process variants (intended).
Variation point — a branching that cannot be reduced to explicit decision statements and thus reflects accidental under-specification (unintended).

In agentic AI, unintended variability arises from non-deterministic LLM generation and from loose-ends in natural-language specifications ("vibe coding"). Observability surfaces these points for the developer to tighten.

Contrast with the software-product-line notion of variability (Czarnecki, Helsen, Eisenecker 2004) which treats variability as desired flexibility for alternative instantiations.

concepts/agent-process-observability · concepts/agentic-bpm · concepts/process-discovery