What is the role of IoT in Agentic AI?

What you need to know

Fast loop versus slow loop

The architecture

1. Devices — stream telemetry (temperature, vibration, location, power).
2. Edge — filter, aggregate, detect anomalies fast; run safety logic locally.
3. Platform — store time series; raise events on anomalies.
4. Agent — on an event, query history, manuals and past work orders; diagnose; propose action.
5. Actuator tool — typed, permissioned commands with limits; human approval for anything safety-related.

Constraints

• Data volume. One sensor at 1 reading per second produces 86,400 readings a day. Never put raw streams into context; give the agent summary statistics and a query tool.
• Connectivity. Links drop. The edge must stay safe without the cloud.
• Security. An agent with actuator access is a physical-safety risk: least privilege, signed commands, rate limits on actuation, and approval for safety-relevant changes.

A real-life example

A dairy company runs cold-chain monitoring for 400 refrigerated trucks and 30 chilling centres.

• Edge: each truck's controller keeps the temperature at 2 to 4 °C and alarms locally if it passes 6 °C. The agent cannot change this.
• Agent trigger: truck 117's compressor duty cycle has risen from 45% to 80% over three days while temperature still looks fine.
• Agent steps: queries the truck's history, the compressor manual and past work orders. It finds two similar cases where a failing condenser fan was the cause. It checks the route: tomorrow's run is 9 hours in 38 °C heat.
• Action: creates a maintenance work order (autonomous), proposes swapping truck 117 off the long route (needs the fleet manager's approval), and notifies the depot.

The fan was failing. Catching it early avoided a likely spoiled load worth about Rs 6 lakh. The agent never touched the temperature controls; its value was in connecting weak signals with history.

Follow-up questions to expect

• "Could the agent adjust setpoints directly?" — Only within narrow, pre-approved ranges, rate-limited and logged, and never bypassing interlocks. Anything beyond that needs a human.
• "How do you feed sensor data to an LLM?" — As summaries and trends (min, max, slope, anomalies), with a tool to query raw history when needed.
• "What happens when the cloud is unreachable?" — The edge runs on its own safely; the agent catches up when the link returns.