How a Delfers agent closes the loop on the plant floor.
This is a technical walkthrough of an autonomous agent operating inside Nexa Edge. It ingests live OT telemetry through the Unified Namespace, reasons over it against learned models and operating constraints, and executes governed closed-loop actions — observe, decide, act, verify — with a human-in-the-loop guardrail on every write. Select a scenario and step through a single decision cycle.
Plant Floor — Live Telemetry
STREAMING · UNSDelfers Agent
IDLELoop closed
What this runs on in production.
The demo above is illustrative, but the architecture underneath it is real. Delfers agents are deployed on the stack below — a vendor-agnostic OT data backbone, edge inference, and a governed actuation layer with audit and human-in-the-loop controls at every step.
| Layer | Technologies | Role in the agent loop |
|---|---|---|
| 01Acquisition | OPC-UAModbus TCPEtherNet/IPHARTDNP3 | Connects to PLCs, SCADA, DCS and field sensors across mixed-vendor estates (Siemens, Rockwell, Schneider, legacy). 250+ industrial protocols, vendor-agnostic. |
| 02Normalisation | MQTT / Sparkplug BUnified NamespaceOPC-UA models | The edge gateway maps OPC-UA nodes into a Unified Namespace and publishes contextualised, event-driven data over MQTT — preserving metadata so the agent reads motor_temp, not an unlabelled tag. |
| 03Streaming & Historian | Apache KafkaTime-series DBStream processing | High-throughput event backbone with persistent replay. Time-series historian retains the signal history the agent's models learn from and reason against. |
| 04Edge Inference | NVIDIA JetsonONNX RuntimeAnomaly / CV models | Sub-10ms inference at the edge — anomaly detection, defect classification, forecasting — running on-prem so raw OT data never leaves the plant network. |
| 05Agent Runtime | OrchestratorPolicy engineGuardrailsLLM reasoning | Multi-agent orchestration: specialised agents observe, plan and decide against operating envelopes and business rules. Every proposed write passes a guardrail and safe-envelope check. |
| 06Actuation | PLC writesCMMS / EAMBMS | Governed closed-loop actions: setpoint adjustments back to the PLC, auto-raised work orders in the CMMS, and building/utility setpoints — with human-in-the-loop approval where policy requires. |
| 07Audit & Visibility | Immutable audit logDigital TwinPower BIESG ledger | Every observation, decision and action is timestamped and immutable. The digital twin and dashboards give leadership live OEE, energy and ESG visibility with a full decision trail. |
See it on your own line in 10 days.
A Delfers pilot puts a live agent on one of your assets in ten days — reading your real telemetry, inside your network, with your guardrails. Pilot to full-factory in four months on the same Nexa backbone.
This page is an illustrative simulation provided for demonstration. Telemetry values, asset identifiers, work-order numbers and event timings are synthetic and do not represent any specific customer, site or deployment. Production agent behaviour, model selection, thresholds, actuation scope and human-in-the-loop policies are configured per site and per asset. Protocol, vendor and product names are referenced for architectural context and remain the property of their respective owners.