Agentic AI for SD-WAN operations

Your NOC, working with a crew of AI agents.

NetPulse watches the telemetry and decides what matters. CEEZ Crew investigates, explains and rehearses the fix. Nothing changes until you approve it, at the level you set.

  • 8agents in NetPulse
  • 24CEEZ Crew workers
  • 14source types read
  • 4levels of autonomy
BR-BER-001 · MPLS primaryExample incident · priority P2
Replay
Latency, path in use20 → 65 ms
AvailabilityUp
Voice quality4.3 MOS
Latency on the path in useacceptable rangeVoice quality (MOS)acceptable limit: 3.6availability: up the whole time+1 h+2 h+3 h
  1. Signal
  2. Understand
  3. Correlate
  4. Decide
  5. Investigate
  6. You approve
  7. Recovered
Signal

Latency climbs from 20 to 65 ms over about three hours

Detection Agent · availability never flips
Understand

Degraded for its hour of the week, with nothing over a fixed line

Anomaly Agent · learns what normal looks like for that hour
Correlate

The link, its tunnel and two applications become one situation

Correlation Agent · symptoms inhibited
Decide

Suspected root: the provider’s middle mile, not the branch edge

Root-Cause Engine · applications are never the root
Investigate

The Transport specialist confirms it. Closed in 2.9 minutes

CEEZ Crew · every claim cites its query
You approve

Steer voice and video to the Internet path until MPLS recovers

Remediation Agent · approval first, rollback attached
Recovered

Latency and voice quality are back inside their limits, verified against the alarm that fired

Verification Agent · the carrier fixes MPLS, then traffic fails back

Example: a silent brownout at a Berlin branch, from first signal to verified recovery.

It reads what you already collect

14 kinds of source become one set of facts and one entity graph.

  • SNMP and traps
  • gNMI
  • gRPC dial-out telemetry
  • NETCONF
  • OpenTelemetry (metrics, logs, traces)
  • NetFlow and IPFIX
  • Syslog
  • Kafka and message buses
  • REST and API polling
  • Prometheus
  • TWAMP and path probes
  • Cloud flow logs and metrics
  • Inventory and CMDB
  • Log shippers

From signal to sign-off

One incident, five hand-offs.

Each stage has an owner and each hand-off is on the record. The first three follow fixed rules. The Crew reasons. A person approves.

  1. 1

    Signal

    Fourteen kinds of source become one picture of your network.

    • Ingest
    • Normalize
    • Entity graph
  2. 2

    Understand

    Breaches are scored against each entity’s own baseline, and repeats are folded.

    • Detection
    • Noise
    • Anomaly
  3. 3

    Decide

    Related alerts become one situation, and a root cause is named, or handed to a person when it is unclear.

    • Correlation
    • Root cause
    • Incident
  4. 4

    Investigate

    Specialists query the data, a commander adjudicates, and every claim cites its query.

    • Triage
    • Specialists
    • Commander
  5. 5

    Approve

    A gated fix with a rollback plan, checked against the alarm that fired.

    • Plan
    • Dry run
    • Verify

From alerts to pages

The last 24 hours.

Raw alerts
277
De-duplicated
83
Firing
70
Situations
12
Pages
11

Each stage removes what another stage already explains: repeats, then flaps, then symptoms of a root.

The graph it walks

How a root cause is chosen, on an example Berlin brownout.

  • Healthy
  • Breaching
  • Suspected root
PROVIDERLINKSTUNNELSAPPLICATIONSSP005Deutsche TelekomBR-BER-001 MPLSprimary · 20 → 65 msInternet backupBR-BER-001 · healthyTUN-4F18D143carries Teams, ZoomBackup-path tunnelhealthyMicrosoft TeamsAPP001 · MOS 3.5ZoomAPP002 · MOS 3.5Other applicationshealthyROOT CAUSENot the cause:1 link affected
  • ProviderSP005 · Deutsche Telekom · not the cause, one link affected
  • LinkBR-BER-001 MPLS primary · 20 → 65 ms · root cause
  • LinkInternet backup · healthy
  • TunnelTUN-4F18D143 · degraded · carries Teams and Zoom
  • ApplicationsTeams (APP001) and Zoom (APP002) · symptoms, never the root
  1. 1Breaches found
  2. 2Attribute upstream
  3. 3Check the provider
  4. 4Conclude
Breaches found

Two applications, their tunnel and one link are all outside their own baselines.

Attribute upstream

The tunnel’s breach is traced to the link underneath it. Applications are evidence, never the root.

Check the provider

A provider is only a candidate when three or more links degrade across two or more sites. Here one link did.

Conclude

The MPLS link explains the tunnel and both applications below it, and nothing above it is breaching. It is the root cause.

96%Fewer pages

277 raw alerts became 11 pages in 24 hours

How it was measured →
77%Root-cause calls correct

17 of 22 calls; unclear cases go to a person

How it was measured →
57 minMedian warning

2 of 5 gradual failures flagged before they breached

How it was measured →
0Fixes left without a rollback

33 governed runs, every one verified

How it was measured →

Results vary with your network, data sources and traffic patterns.

The agents

Named for the job, not the model.

32 agents in four groups. Each one lists what it reads, what it produces and how much it may do on its own.

8 agents

NetPulse agents

8 agents that detect, group, name a root cause and file the incident. Rules and statistics rather than a language model, so every call is reproducible.

77%root-cause calls correct
See these agents
17 agents

NOC crew

17 workers take an incident from triage to a verified close. Specialists query the data and a commander adjudicates.

2.9 mindetection to close
See these agents
3 agents

Assurance squad

3 workers on a schedule rather than an incident: baselines, forecasts and maintenance windows.

57 minmedian warning
See these agents
4 agents

Governance squad

4 workers that turn telemetry into SLA claims, vendor scorecards, config audits and executive briefs.

4report types
See these agents

Use cases

Start from what you’re seeing.

Pick the situation. Each use case is a short procedure: how to recognize it, which agents are already on it, what you do, and what the answer means.

Something is down

A circuit, a region or a device has stopped. You need the cause and the next move.

First: A circuit went down

  1. Confirm what is actually downAlerts
  2. Check that traffic really movedSituations
  3. Size the damage on backupExplorer
  4. Open the carrier ticket from the incidentGoverned Brain

Measured outcomes

Outcomes you can check.

Fewer pages, correct root causes and earlier warning, each with how it was measured.

96%

fewer pages than raw alerts

Alerts to pages

24-hour test window

277 raw alerts became 11 pages through de-duplication, correlation and suppression of symptoms.

How it was measured →

77%

of calls correct

Root-cause calls

29 benchmark incidents

17 of 22 calls. Where the evidence is unclear it says so and hands over to a person.

How it was measured →

57 min

median warning

Early warning

28 benchmark incidents, gradual failures

2 of the 5 gradual-onset failures were flagged before they breached.

How it was measured →

Worked examples

Three real events, start to finish.

Berlin · brownout

A silent brownout at a branch

What happened
Latency climbed from 20 to 65 ms over about three hours while the link never reported down.
What the agents did
The Anomaly Agent flagged it against its own baseline. The Transport specialist isolated the provider’s middle mile.
Result
Closed in 2.9 minutes, against a 120-minute runbook target.
Follow this use case
Chicago · hard-down

A circuit goes down

What happened
The MPLS primary dropped. The tunnel failed over to a metered 4G backup in about 28 seconds.
What the agents did
Detection fired, the Incident Agent filed one ticket, and the Remediation Agent proposed the failback.
Result
Failback waits for the primary’s availability to hold for 15 minutes, so a flap does not bounce traffic.
Follow this use case
Berlin · security

A flood of intrusion alerts

What happened
About 5,000 IPS hits after a ruleset update: nearly all low-severity from internal addresses, three High from outside.
What the agents did
The Security specialist set the internal flood aside so the three that mattered surfaced.
Result
That fix has been verified 8 times, so it now runs automatically with a notification instead of waiting for approval.
Follow this use case

Control

Autonomy is earned.

Agents start by suggesting. A fix moves up a level only on evidence, and moves back down on a regression. The level is set per playbook and per kind of incident.

  1. 1

    Suggest only

    The agent proposes and explains. People do the work.

    Recommend only
  2. 2

    Approve every action

    A named person approves each step before it runs.

    Approval first
  3. 3

    Approve the exceptions

    Routine work inside the limits runs on its own, and tells you.

    Automatic with notice, once earned
  4. 4

    Autonomous within limits

    Runs inside a bounded policy envelope, on the record, reverting on a regression.

    Within a set envelope
  • Recommend only is where every playbook starts, and where one stays until it has a proven record and a rollback plan.
  • Approval first needs a good record for that kind of incident. Automatic with notice needs a strong record, repeated verified recoveries and a proven rollback.
  • A regression reverts it. The demotion is recorded.

CEEZ Crew

Give a crew your own objective.

A worker has a role, instructions, constraints, a model and an autonomy tier. Tools are granted step by step and every run leaves a record. NetPulse hands incidents to whichever team you point it at.

Transport SpecialistCrew specialist
Role
Tell a provider brownout from congestion from a last-mile flap
Reads
Link performance grouped by provider and siteOne tool: read-only query
Constraints
Cite the query behind every claimReads evidence, never assumes
Model
Your model connectionNothing is hard-wired
Autonomy tier
MonitoredRead-only in this team
Hands over
A verdict with link codes and numbersTo the commander

Questions

Asked before anyone starts.

What is the difference between NetPulse and CEEZ Crew?

NetPulse is the set of agents that watch telemetry, group alerts, name a root cause, file incidents and gate fixes. They use rules and statistics, so every call is reproducible. CEEZ Crew is a workforce of model-backed workers that NetPulse hands an incident to: specialists that query the data, a commander that adjudicates, and writers that explain.

Does it change my network?

Only what you approve. Every fix is a proposal with its blast radius, its rollback and the way recovery will be verified. You approve it, hold it, or take the recommendation into your own change process.

Where do the numbers come from?

From a 24-hour window and a set of reference incidents with known causes. The outcomes page shows each figure with what it counts and what to keep in mind. Your own results are on the Accuracy screen once it is running on your network.

What happens when it is not sure?

The incident goes to a person with the evidence gathered so far, rather than a guess. The decision is logged, so you can see how often it happens and why.

Do I need a language model?

Not for detection, correlation, noise handling, forecasting or root-cause analysis. The Crew workers use a model connection you provide. Automatic investigations are off by default and are capped per day and by spend.

Which data sources does it take?

SNMP and traps, gNMI, gRPC dial-out telemetry, NETCONF, OpenTelemetry, NetFlow and IPFIX, syslog, Kafka and message buses, REST polling, Prometheus, TWAMP and path probes, cloud flow logs and metrics, inventory and CMDB, and log shippers.

Something is wrong right now?

Open Situations and pick the use case that matches what you see.