Illustrative reference architecture. Validate requirements for your organization.
Illustrative architecture: A water utility receives low-pressure alarms and customer reports across several neighborhoods. Operators need to reconcile telemetry, planned works, pipe topology, and field observations before dispatching crews or changing network configuration. Agents prepare situational evidence without operating pumps, valves, or public-health decisions.
Problem
Customer reports may be duplicated or geocoded incorrectly, while telemetry can fail during the same incident. Inferring a pipe break from a pressure pattern alone can misdirect crews. Network topology and isolation boundaries require engineering validation.
Why a multi-agent system?
Telemetry, asset-graph, field-work, and customer-impact specialists investigate different evidence channels. The workflow keeps confirmed observations separate from modeled exposure and routes consequential decisions to established utility roles.
Intended outcome
The intended result is a reviewable operational decision with source evidence and explicit ownership. Assess evidence quality and coordination time alongside field response. Service restoration and public-health outcomes require accountable operational decisions; the agent contribution is an inspectable, current picture of what is known and what still needs verification.
Specialist agent team
Telemetry investigator
Compare pressure and flow observations with sensor health and planned operating changes.
Tools
Historian read API
sensor diagnostics
Authority boundary
Cannot set pump targets or change alarm thresholds.
Asset dependency agent
Trace candidate pipe segments and connected service areas using governed network data.
Tools
Geographic information query
asset graph
Authority boundary
Cannot authorize isolation boundaries or assume map accuracy.
Field coordination agent
Locate active work orders, crew availability, and verified field reports.
Tools
Maintenance read API
crew status query
Authority boundary
Cannot dispatch unsafe work or bypass site access procedures.
Customer impact agent
Group reports and draft factual service updates from approved incident status.
Tools
Customer case search
notification draft tool
Authority boundary
Cannot issue health guidance or publish notices without authorized review.
Enterprise systems
Operational telemetry gateway
Exports observations without providing a write path to control equipment.
Geographic and asset management systems
Provide pipe topology, valves, maintenance history, and map revision metadata.
Incident and customer platforms
Own crew assignments, approved public statements, and report follow-up responsibilities.
Architecture diagrams
Water network incident coordination — logical architecture — Specialists produce bounded findings; the evidence ledger preserves source versions and human decisions before an authorized adapter changes a system of record.
Failure and decision sequence — A pressure sensor fails during the incident. The workflow records unresolved work and resumes only after the source state and reviewer authority are checked.
sequenceDiagram
participant O as "Case orchestrator"
participant S as "Operational telemetry gateway"
participant H as "Accountable reviewer"
O->>S: Read pressure and asset context
O->>H: Present possible affected segments
H-->>O: Request field verification
S-->>O: Field topology contradicts map
O->>H: Suspend map-dependent response proposal
Runtime flow
1
Correlate incoming reports
Deduplicate customer cases and alarms by location and time while retaining individual evidence and uncertain addresses.
2
Check telemetry reliability
Compare sensor health, neighboring observations, and recent operational changes before accepting a pressure pattern.
3
Map possible exposure
Trace affected segments using current asset data and label modeled impact separately from confirmed service disruption.
4
Gather field context
Retrieve active works and verified crew observations; surface conflicts between map records and physical findings.
5
Prepare operator options
Show inspection priorities, uncertain isolation assumptions, and a draft customer update for the appropriate operational owner.
6
Track authorized response
Record approved dispatch and communication receipts, then reconcile field findings and telemetry before closing the incident.
Failure recovery
A pressure sensor fails during the incident.
Detect: Sensor diagnostics and neighboring measurements disagree with the isolated reading.
Recover: Exclude the suspect signal from confirmed evidence and request field validation without suppressing the incident.
A field crew discovers a valve absent from the map.
Detect: Verified field report conflicts with the asset revision used in the plan.
Recover: Suspend map-dependent recommendations and route topology correction to engineering before further isolation planning.
Security & decision boundaries
Prevent agent access to pump, valve, and supervisory-control write interfaces.
Require operational approval for dispatch and separate authorized review for public-health messages.
Label confirmed outages, customer reports, and modeled exposure as distinct evidence classes.
Protect customer addresses in aggregated incident views and telemetry exports.
Attach asset-map revision and field verification status to every network-dependent proposal.
Operational metrics
Measurement definitions, not promised improvements.
Situational packet latency
Time from incident registration to operator-reviewable evidence and proposed next checks.
Location resolution quality
Audited reports assigned to the correct service area divided by audited geocoded reports.
Field contradiction rate
Plans requiring revision after verified field findings divided by plans reviewed in the field.
Communication consistency
Published updates matching approved incident state divided by sampled published updates.
Architecture tradeoffs
Decision
Benefit
Cost
Multi-source correlation
Reduces reliance on any single failed sensor.
Requires careful uncertainty labels and location reconciliation.