Contextual Agentic

A Factory Is Not an API Sandbox

Manufacturing agents can analyze and propose, but physical commands require a separate safety and operations boundary that can refuse or stop the action.

·5 min read

Written as at 22 October 2026.

industrial-aiot-securityroboticsagentic-aisafety

A software agent can make a bad API call and leave a record that helps us reverse it. A physical system can crush a component, stop a production line or put a person in danger before the record is written. That changes what we should let an agent do.

At the illustrative MAAH Group factory, a maintenance agent notices abnormal vibration on a robotic cell. It can correlate the sensor trend with the work-order history and propose an inspection. That is useful. But if the same agent can remotely change the cell’s speed limit or restart it after an interlock, the architecture has crossed from analysis into physical control.

Two boundaries, not one

The first boundary is the enterprise agent gateway. It authenticates the human, agent and task; limits what data may be read; validates a typed proposal; and records the evidence. The second is the plant’s operational and safety boundary. It knows the machine state, local permits, maintenance isolation, physical interlocks and whether the proposed command is permissible now.

An agent should not be able to redefine either boundary by writing a new prompt, calling another tool or finding a more permissive MCP server. When the plant controller refuses, the agent’s recovery path is to report and escalate, not to search for an alternate route.

Agent output Safer receiving path
“The vibration trend merits inspection” Create a draft work order with links to readings and prior maintenance.
“Pause the cell” Route to a named operator or a pre-approved local automation with bounded conditions.
“Restart the cell” Require plant-state checks and the safety/operations procedure; do not use a generic restart tool.
“The interlock is faulty” Open an engineering investigation; the agent cannot suppress or bypass the interlock.

The physical safety controller must remain effective when the model, network or orchestrator fails. An agent may improve the speed of diagnosis, but it cannot be the final authority on a safety-relevant state transition. Keep execution credentials outside the model and narrow the tool contract to the allowed procedure.

The test to run before deployment

Simulate stale telemetry, a forged maintenance note, a missing operator approval and a plant network partition. For each, ask whether the proposed action is refused at the last authoritative boundary. Then examine the evidence record: can an investigator see the sensor inputs, model and tool versions, proposal, policy decision, human approval and controller response as separate events?

That test is harder than a dashboard demo, and more useful. It shows whether an agent can assist a factory without becoming an undocumented route around its operational controls.

SAE develops the containment and physical-world design in detail. See the books page for the edition’s status.

References

  1. The Contextual Secure Agentic Enterprise, first edition, Chapters 18 and 22 (prepublication manuscript, 22 October 2026).

← All articles

Related