Architecting Network Automation with Software Engineering Principles#

Leandro Lerena · zebbra AG (per slides)

• Why: The most genuinely novel technical idea in the program: applying database-transaction semantics and typed safety profiles to network operations — and deriving a principled safety model for agentic AI from it.
• What: A response to scripts reaching their structural limits (no versioning, unknown safety profiles, undetectable race conditions, weak auditability). The replacement: "function blocks" with declared properties (pure / idempotent / reversible), pessimistic locking, and a workflow state machine distinguishing FAILED_SAFE (pure steps only, or rollback succeeded) from FAILED_UNSAFE (side effects may have occurred; resources stay locked to prevent more workflows from touching unknown state).
• How: Race conditions are treated as the read-modify-write problem relational databases already solved — a central orchestrator locks context, executes blocks transactionally, rolls back reversible ones, and escalates unsafe ones. Vendor plugins (Cisco IOS via Scrapli, SR Linux via Netmiko, etc.) implement capability interfaces so business logic stays vendor-agnostic; every block is tested against netlab virtual labs via pytest fixtures (remote-lab developer preview on GitHub). The agentic tiers fall out naturally: autonomous read-only agents (pure blocks only), autonomous allow-list (pre-approved idempotent/reversible blocks), and supervised everything else — agents never touch device credentials.
• Where applicable: Greenfield framework thinking from a consultancy, not a hyperscaler postmortem — but the safety-tier model is directly reusable by anyone designing the agent-to-network boundary, which this PM will likely have to opine on soon.