Devices & Ecosystems
An agent embedded in hardware doesn't just execute tasks — it builds a causal model of its physical environment. Sensors are signals. States accumulate. Patterns emerge. And humans can audit, correct, and extend what the device has learned, in plain language, any time.
This lane is forward-looking. The core tools work today. The full on-device deployment vision — agents that learn locally on embedded hardware with world models syncing to the edge — is in active development. Build with this in mind.
Overview
Traditional IoT architectures push data to the cloud and pull decisions back down. An agent with a persistent world model changes this: the intelligence lives on the device, the model accumulates locally, and humans interact with it in natural language — without querying dashboards or writing SQL.
Every sensor reading is a potential signal. Every state change is an event. Every recurring behavior is a pattern. Nervous Machine gives your device agent the scaffolding to learn from its environment, accumulate this knowledge over time, and act on it with growing confidence.
The Device Model
A device world model captures the physical environment the device operates in — not just what the device measures, but what those measurements mean.
| What | Signal type | Example |
|---|---|---|
| Sensor baselines | metric | Ambient temp normal range: 18–24°C |
| Current device state | state | HVAC: cooling / idle / fault |
| Recurring behaviors | pattern | Temp spikes 2°C every weekday at 9am |
| Environmental claims | claim | Room B air quality degrades when west window is open |
| Incidents & anomalies | event | Power spike on 2025-11-03, cause: unknown |
| Connected devices | entity | Sensor-node-4, gateway-main, thermostat-A |
Sensors as Signals
Each sensor reading can be saved as a metric event. Over time, the world model builds a statistical baseline for each sensor — and deviations from that baseline become meaningful signals in themselves.
Normalize sensor values to 0–1. Map the expected range of a sensor to [0, 1]. A temperature sensor that normally reads 18–24°C would map 22°C to ~0.67. This lets the world model compare signals across different sensor types on a common scale.
States & Patterns
Current state
Use state events to track what mode or condition the device or system is currently in. Update these as transitions occur — the world model accumulates a history of state changes over time.
Patterns
When the agent detects a recurring behavior — confirmed across multiple observations — save it as a pattern event. Patterns let the agent anticipate and act pre-emptively.
Anomaly Detection
When a sensor reading significantly deviates from the established baseline, save it as an event with low certainty (cause unknown) and flag it for human review via a curiosity trigger.
Human Alignment
Devices observe physical reality — but humans provide intent, context, and causality that no sensor can supply. Why was the window left open? Was the spike a test or a fault? What should the device prioritize when it conflicts?
Ecosystem Context
Devices don't operate in isolation. Use link_events to connect related signals across your device ecosystem — a temperature anomaly linked to a power spike, a state change linked to a known pattern.
Signal Types Used in This Lane
| Signal | Use case |
|---|---|
| metric | Sensor readings, environmental measurements |
| state | Device mode, operational status, phase |
| pattern | Recurring behaviors, environmental cycles |
| event | Anomalies, incidents, state transitions |
| claim | Environmental rules, structural constraints |
| entity | Sensors, gateways, connected devices |
| preference | Device priorities, human-set intent |