Agent Process Observability

The discipline of inferring actionable insight into the inner workings of an Agentic AI system by analyzing the inputs, outputs, and execution trajectories as they flow between agents (sources/2025-fournier-agentic-ai-process-observability). Borrows terminology from the Agentops line of work (Dong et al. 2024, (referenced-not-ingested)).

Core idea

Treat each timestamped tool invocation by an agent as an event. A run of the agent system becomes a trace; a batch of runs becomes a process event log amenable to concepts/process-discovery and causal process discovery.

Key distinction

Intended variability — decision points the developer explicitly specified.
Unintended variability — accidental variation points emerging from under-specified prompts, tools, or roles.

Distinguishing the two is the main diagnostic task. Fournier et al. attack it with LLM-based static analysis, matching a derived gateway rule statement against the agent specification source text.

Why it matters

LLM-backed agents are non-deterministic even with temperature=0 (Ouyang et al. 2025). Without dedicated observability, silent breaches of responsibility, loophole tool usage, and specification drift go undetected.

concepts/agentic-bpm · concepts/behavioral-variability · concepts/causal-process-discovery · concepts/explainability-apm · concepts/self-modification