Agentic Network Operations (NetOps) is emerging as a paradigm that enables workload‑aware, self‑adjusting and reliable autonomous networks. While agents have shown value in incident summarization and telemetry extraction, their role as autonomous control‑loop engines still hinges on long‑horizon reliability.
In a datacenter fabric, an agent must react to alarms and operator intents while abstaining from high‑risk actions that could cause or extend downtime. Abstention presupposes that the impact of an action is known before execution, yet existing NetOps benchmarks lack per‑action ground truth.
We therefore construct an action‑level ground truth for the network repair task in NetArena. By symbolically replaying the emulated network and validating against the environment at every step, we obtain the exact value of each action.
From this value we derive two pre‑execution targets: whether an action reduces the repair distance (progress) and whether it increases it (harm).
Across ten agent models, verifiers that leverage internal signals predict both progress and harm more reliably than a baseline that uses observable signals only.
Looking ahead, these safety signals can serve as feedback to an agent harness, enabling it to abstain from risky actions and protect the target system.
Review