Manufacturing agents need machine context before machine control.

Interoperability is a foundation, not a diagnosis

The OPC Foundation describes OPC UA as a platform-independent architecture with information modelling and access-controlled reads, writes, subscriptions and methods. Those capabilities can expose equipment context, but do not by themselves interpret a fault or prescribe a maintenance action.[1]

Documents can include more than plain text

DeepSeek-VL’s 2024 paper describes a vision-language model using varied data including PDFs, charts and web screenshots. This research motivates investigating visual document understanding. It does not validate the model on a particular factory manual, engineering drawing or inspection workflow.[2]

Propose an incident evidence packet

Choose one line and one recurring class of incident. A read-only assistant could collect the alarm timestamp, asset identity, permitted signal history, relevant manual sections and recent work orders. Keep observations separate from inferred explanations. Show the exact manual revision rather than merging conflicting instructions.

Map equipment identifiers across operational and maintenance systems before adding language reasoning. A temperature value without units, time or sensor identity is not sufficient evidence. The system should surface missing or stale data and propose the next permitted information-gathering step.

Keep operational authority with the existing system

Draft a technician handoff with evidence, uncertainty and an approved checklist reference. Do not let the model generate novel machine settings or bypass a lockout procedure. Any production control integration is a separate engineering project requiring the plant’s documented safety and change-management process.

For the first pilot, expose only read tools and a draft work-order interface. Record who reviewed the proposal and which records were used. A new alarm or changed machine state should invalidate conclusions that depended on the older snapshot. This is a deployment concept, not a claim of reduced downtime.

Test retrieval and judgement independently

Evaluate whether the assistant retrieves the right asset, time range and manual version. Then have qualified technicians assess whether the draft distinguishes observations from hypotheses and follows the approved escalation path. Include sensor failures, conflicting documentation and incomplete histories.

Measure investigation effort, missed context, unsupported explanations and reviewer corrections. Assess the workflow alongside existing incident handling. A useful assistant should make the evidence easier to inspect; the existence of an interoperable protocol or a vision model does not establish fault-detection accuracy.