Six agents, each owning a layer of network operations. Not a feature bolted onto a monitoring tool — the agents are the architecture.
What does an AI agent actually do in a NOC?
Most “AI” in network operations is a summarization layer: it reads what your monitoring stack already produced and writes a paragraph about it. That saves reading time and nothing else. The work — correlating, diagnosing, answering the customer, updating the ticket, briefing the supervisor — still lands on a person.
A Rapax agent owns a job end to end. It has a defined mandate, it acts inside that mandate without waiting to be asked, and it hands off with context attached when the job needs a human. The six below cover the operational surface where headcount currently scales with network size.
What does each agent actually do?
Each card below is the same shape: what sets the agent off, what it does about it, and what you get. The demo link on each one shows it running against a live FTTH network.
Frank
- TriggerA ticket arrives by email, portal, alert, or API
- Frank doesDeflects what Wade can answer, classifies the rest, requests missing logs, checks reported outages against live telemetry
- ResultRoughly 40% deflected and 15% auto-closed; the rest escalate with context already attached
Oscar
- TriggerSomeone asks what is happening — or a shift changes
- Oscar doesMaintains a continuous operational picture: open incidents, customer impact, SLA exposure, what is waiting on a human
- ResultA briefing, not a dashboard — down to the affected service and the PON port
Wade
- TriggerAnother agent, or an engineer, asks how to fix something
- Wade doesRetrieves the runbook, procedure, or vendor document — and returns a confidence level with it
- ResultGrounded answers. If the procedure is not documented, Wade says so rather than improvising
Grace
- TriggerA service event is confirmed, or a stakeholder needs an update
- Grace doesComposes and sends notifications on the right channel, respecting subscriptions and quiet hours, and tracks delivery
- ResultNobody stops working the fault in order to tell people about the fault
Nora
- TriggerAny question about what is connected to what, or who is affected
- Nora doesAnswers from live topology and as-built documentation together, and flags where the two disagree
- ResultOne source of truth, queryable in plain language — and the foundation correlation depends on
Bruce
- TriggerA new device type, vendor, or external system needs connecting
- Bruce doesBuilds the integration against the vendor SDK — Syslog, SNMP, webhooks, REST — with tests, and commits the code for review
- ResultAdding a vendor is a configuration task, not a professional services engagement
How do they work together on one incident?
An OLT line card fails. Every ONT behind it loses GPON signal and reports independently — forty-nine alarms describing one event.
Correlation collapses them into a single service record with root cause, affected segment, subscriber count, and service tier. Nora supplies the topology that makes the collapse possible. Grace notifies the affected subscribers and internal stakeholders from that one record. Frank handles the support tickets that arrive anyway, answering from Wade rather than adding to the queue. Oscar’s briefing reflects all of it continuously, so the supervisor walking in at shift change reads one page instead of reconstructing the night.
No agent in that sequence is doing something a person could not do. The difference is that all of it happens at once, and none of it waits for someone to notice.
What stops them from making things up?
The failure mode of AI in operations is well documented: confident, wrong, and fast. Rapax addresses it structurally rather than with a better prompt.
Every agent decision is anchored to Wade, which returns a confidence label on every query. You set the thresholds that map confidence to action — high confidence acts, medium drafts for review, low escalates. If the answer is not in your documentation, no agent invents one. And priority changes that carry billing or paging consequences stay with the human.
Rapax also runs on any LLM provider, including fully local deployment. If your topology and subscriber data cannot leave your environment, they do not have to.
See them running on your own network
Rapax deploys in five days. You evaluate for thirty against success criteria you define in writing before anything is signed, for $25,000 — credited against the license if you move forward. If it does not meet them, the evaluation ends and you owe nothing further.
The ask is fifteen minutes, not a demo. Book 15 minutes · sales@rapax.app · Read the full Frank writeup