Forward Deployed: Agents in the NOC

The first time I saw a fiber cut, it was physical.

Back in the Alcatel-Lucent NOC in Plano, we monitored dark fiber for AT&T, Verizon, Time Warner. When a backhoe in Ohio severed a 144-strand cable, it showed up on our screens as a red line on a map—a sudden loss of light. The 1640 and 1670 nodes would stop responding, OTDR traces would show the exact distance to the fault (light travels one foot per nanosecond; the reflection tells you exactly where the glass broke), and we’d dispatch a truck.

The NOC was a room full of humans staring at dashboards, correlating alarms, making phone calls. Human latency measured in minutes, sometimes hours. The network could heal itself in milliseconds if it knew what to do, but it didn’t. It needed us to interpret the light.

Fifteen years later, I’m building agents that do what we did—but with a constraint we never had: tokens cost money.

The Physics Have Changed

In 2008, the bottleneck was human attention. One engineer could watch maybe twelve circuits before the cognitive load became dangerous. Today, the bottleneck is computational economics. An agent watching twelve thousand circuits isn’t limited by attention—it’s limited by the cost of inference.

This is why “AI transformation” fails in infrastructure. Companies buy Claude Enterprise seats, plug in their network data, and watch the token meter spin like a gas pump. They treat agents like employees (salary + benefits) when they should treat them like capital equipment (upfront cost + marginal operational expense).

Forward Deployed Means Privileged

The term “forward deployed” comes from defense contracting—engineers who work inside the customer’s secure perimeter, not from a vendor’s office. In the NOC context, it means agents that operate on privileged service accounts with access to live network state, not sanitized datasets exported to a chat window.

This creates a paradox: the more access you give an agent, the more useful it becomes, but the more catastrophic a prompt injection or hallucination becomes. A forward-deployed agent with write access to a Juniper router configuration isn’t a chatbot—it’s a potential outage.

The solution isn’t less access. It’s audited access.

Every prompt executed in a production NOC environment needs:

• Immutable logging: Input hash, output hash, token count, execution time, model version

• Intent interception: Guardrails that score “should this agent be able to reboot that node?” before the tool call executes

• Human-in-the-loop for destructive ops: Agents can diagnose; humans commit changes

The CPNI Constraint

Telecom data isn’t just sensitive—it’s regulated. Customer Proprietary Network Information includes call detail records, network usage patterns, even the physical location of fiber infrastructure. When I was monitoring those networks, we couldn’t email raw CPNI to external consultants without legal review.

Today, sending that same data to an external LLM API is a compliance violation waiting to happen. Forward-deployed agents must run on local inference—models that live inside your VPC or air-gapped entirely. The prompt library becomes a governance artifact: versioned, reviewed, and deployed like code, not shared in a Google Doc.

The Half-Life of Context

Here’s what I didn’t understand in the NOC: every decision has a decay curve. The alarm comes in, the context is fresh, the technician remembers what they were doing. Wait an hour, and the mental model evaporates.

Agentic systems have the same problem. A forward-deployed agent troubleshooting a fiber cut needs the current OTDR trace, the last 24 hours of performance data, and the physical topology of the route. That’s a lot of context—thousands of tokens. And context has a half-life. Wait too long to act, and the situation changes; act too fast with stale context, and you make the wrong decision.

The best agents aren’t the ones with the biggest context windows. They’re the ones that know when to compress, summarize, and escalate—when to burn tokens on reasoning and when to route to a human.

From Monitoring to Action

The NOC I worked in was reactive. Alarms fired, humans responded. The agents I’m building now are predictive. They watch the same fiber nodes I did, but they notice the 0.1dB degradation trend before it becomes a hard fault. They draft the dispatch ticket, check the parts inventory, and route the truck—while the human supervisor reviews the plan.

But they do it within a token budget. Because in infrastructure, the goal isn’t to have a conversation. It’s to keep the light moving.

One foot per nanosecond. Zero excuses for downtime.