Toward Agentic Network Operations: What AI Can (and Can’t) Do#

Tobias Bühler - netfabric.ai

• Why: Not one of your four core areas, but if agentic identity is on your radar, this is the ops-side mirror: a fact-driven talk (by an ETH Vanbever-group PhD, now NetFabric CTO) with real numbers on where agentic AI in networking actually stands — including the failure modes and the cost math most talks omit.
• What: An honest positioning: most operators sit between awareness and experimentation; production use is blocked by hallucinations, privacy, unstable results, and cost. The concrete worked example is syslog: a ~100-device network produces millions of log lines/day; naive LLM ingestion costs ~$500k/year (100 tokens/log, 5M logs/day, $3/1M tokens) and overflows any context window. Their key mechanism: LL-generated parsers that compress 69M raw log lines into ~700 typed event types — and their deployment data shows new event types keep appearing even after a month (OS upgrades, maintenance days), so parser generation must be continuous. Documented pitfalls: overly specific parsers, overly general parsers, missing domain knowledge. Five enablers that measurably improved performance: in-context learning with clustering-selected diverse examples, injecting expert parsers for known parameters (e.g., MAC-address formats), guardrails/error checks, targeted system prompts, dedicated log-prefix handling. On agentic data access, a clean control-vs-flexibility framing: agent-written SQL queries (expressive, unsafe) vs targeted tool calls (safe, limited) — recommendation is to choose per LLM model, not one-size-fits-all.
• How: Log abstraction layer first (parsers → events), then LLM reasoning over events rather than raw logs; the agent interacts with structured stores via either SQL or curated tool interfaces.
• Where applicable: Brownfield multi-vendor environments; the numbers come from NetFabric's own deployments (vendor talk, but unusually candid about limitations). The parser-generation pattern itself — LLM + clustering + human correction — is directly stealable for any log-heavy system, networking or not.