A measure of how well internal states of a system can be inferred from knowledge of its external outputs.
Observability matters most in black-box AI systems. In those systems it extends past traditional logging to tracing multi-step Agentic Workflows, monitoring GPU thermals and VRAM saturation, and detecting semantic drift in LLM responses. Those signals give teams what they need to debug non-deterministic systems before problems escalate.
See also Agentic Workflows, Air-Gapped Environment, Role-Based Access Control and Zero Trust Architecture.