Architecture discipline + AI guardrails

A fast network is not one giant smart machine. It is more like a restaurant kitchen where one station takes orders, one cooks, and one plates, because the meal slows down when every cook also answers the phone. (infoq.com) That is the point Amir Langer made in an InfoQ talk this week: separate business logic from input and output work, because mixing the two adds delay even before you reach the hardware tricks. InfoQ’s summary says that split is one reason tools like Aeron and the Disruptor can reach single-digit microsecond speeds. (infoq.com) Network automation is arriving at the same idea from the operations side. John Capobianco wrote on February 2, 2026 that serious teams are moving away from one monolithic “artificial intelligence” helper and toward specialized agents that read telemetry, understand routing intent, validate risk, and coordinate remediation. (itential.com) That split matters because each job needs a different kind of evidence. A collector gathers raw signals, an observability layer turns those signals into traces and alerts, an intent layer states the target outcome in policy terms, and an orchestration layer decides the order of actions. (ericsson.com) Ericsson described that architecture on February 24, 2026 as “intent-driven” operations, where humans set measurable goals and bounded domains carry them out. Its paper says autonomy only works when observability, policy bounds, and human oversight are built into the design from the start. (ericsson.com) Nokia pushed the warning further on February 13, 2026. It said isolated artificial intelligence loops in radio, core, transport, and access networks can fight each other, creating oscillations and “priority inversions,” which is engineer-speak for one fix accidentally making another system misbehave. (nokia.com) Its example was congestion control. If one system reroutes traffic, another system may react to the new pattern, then a third system may react again, and the network can end up “chasing its tail” unless one coordinating layer reconciles the objectives. (nokia.com) Now add generative artificial intelligence to that stack and the risk changes shape. An agent can draft a configuration in seconds, but if it also has broad production permissions, it stops being a helper and starts acting like a senior operator with no memory of last month’s outage. (infoq.com) InfoQ’s February 23, 2026 piece on infrastructure automation argued for a hard boundary outside the execution systems themselves. It recommends a dedicated agent gateway, policy checks on every action, OpenTelemetry auditing, and short-lived runners so each change happens in an isolated sandbox with predictable cleanup. (infoq.com) OpenAI’s own Model Spec update on February 12, 2025 made the same governance point in software terms. It says models should follow a clear chain of command and operate inside “well-defined boundaries,” which is another way of saying the model should not be the final authority over sensitive actions. (openai.com) Put those pieces together and the industry message this week was blunt: do not let one giant agent observe, decide, and execute inside production by itself. Split the system into roles, make intent explicit, log every step, and require approvals or policy gates before machine-written changes touch live infrastructure. (infoq.com) (ericsson.com) (infoq.com)