embedded-agent.com

Embedded Agent Glossary: 12 Key Terms Defined

Last reviewed: 2026-05-22 · Marcus Rüb

Glossary of Embedded and Edge AI Agent Terms

Definitions for twelve terms that appear frequently in discussions of embedded AI agents, edge intelligence, and industrial IoT. Each definition is written to be technically precise and distinguishes the term from closely related concepts.

Embedded AI Agent

Software that combines autonomous agent behaviour (goal-directed perception, reasoning, and action) with machine-learning inference, running entirely on constrained edge hardware — such as microcontrollers, industrial controllers, or edge gateways — without requiring cloud connectivity for its core decision loop. The defining properties are local execution, statefulness, goal orientation, and action capability. See Embedded AI Agent for a full treatment.

Edge AI Agent

An autonomous software agent whose sense-decide-act loop executes on hardware located at or near the data source — at the network edge — rather than in a central cloud. “Edge AI agent” is a broader term than “embedded AI agent”: it encompasses agents running on powerful GPU-equipped edge servers as well as MCU-class devices. All embedded agents are edge agents; not all edge agents are embedded agents.

MCU Agent

An autonomous software agent designed to execute on a microcontroller unit (MCU) — a single-chip device integrating CPU, RAM, Flash, and peripherals, typically with memory measured in kilobytes to low megabytes and power consumption measured in milliwatts. MCU agents are highly constrained and rely on rule-based reasoning, small quantized models, or hybrid approaches. Examples of MCU families commonly targeted include STM32, ESP32, nRF52, and TI MSP430/MSPM0 families.

Industrial Edge Agent

An embedded or edge agent deployed in a manufacturing, energy, utilities, or critical infrastructure environment, operating under industrial communication standards (OPC UA, Modbus, PROFINET, Sparkplug B) and typically subject to formal safety requirements (IEC 61508, IEC 62443). Industrial edge agents often operate air-gapped or on isolated OT networks and must coexist with legacy control systems. See Industrial Embedded Agents.

Agentic Edge

A system architecture in which edge devices are not passive data sources but active, autonomous agents — capable of local reasoning, decision-making, and coordination with peer agents, without dependence on a central orchestrator for every action. “Agentic edge” describes the design philosophy that intelligence belongs at the data source, not only in the cloud.

On-Device Agent

Synonym for embedded agent, emphasising that the agent’s processing occurs on the device itself rather than being offloaded to a remote server. The term is common in consumer electronics and IoT product contexts; “embedded agent” is more common in industrial and engineering contexts.

Local AI Agent

An AI agent whose inference and reasoning execute on local hardware — whether an MCU, a gateway, an industrial PC, or a local server — without sending data to cloud infrastructure for the core decision. The term encompasses both MCU-class embedded agents and larger local deployments (on-premise servers, edge data centres). The defining property is data locality: the model and the data stay on-site.

Hybrid Edge-Cloud Agent

An agent architecture in which time-sensitive and privacy-sensitive decisions are made locally on edge hardware, while less time-critical tasks — fleet analytics, model retraining, natural language interfaces, and long-horizon planning — are handled by cloud agents. The two layers communicate asynchronously; the edge agent remains functional when the cloud connection is unavailable. Hybrid architectures represent the most common production pattern for sophisticated embedded agent deployments. See Embedded Agent vs Cloud Agent.

Agent Registry

A directory service that allows embedded agents to announce their presence and capabilities, and enables other agents or orchestrators to discover, monitor, and interact with them. In MQTT-based systems, registries are commonly implemented using retained messages and Last Will and Testament (LWT). EMQX Enterprise 6.2 (2026) introduced a native A2A Registry that implements agent discovery over MQTT following emerging Agent-to-Agent protocol specifications. See Agent Registry for Embedded Systems.

AI Service Technician

An emerging concept in industrial maintenance where an agent — embedded in or accessible from a technician’s handheld device — uses sensor data, local knowledge bases, and AI reasoning to guide the technician through fault diagnosis and repair procedures. The agent operates locally using cached models and documentation, synchronising with central systems when connectivity is available. This pattern is emerging in heavy industry, energy infrastructure, and complex machinery maintenance contexts.

TinyML

A discipline within machine learning focused on compressing, quantizing, and deploying trained models on ultra-resource-constrained devices — microcontrollers and similar hardware with memory measured in kilobytes and power in fractions of a watt. TinyML covers the toolchain (training, quantization, pruning, conversion) and the inference runtimes (TensorFlow Lite for Microcontrollers, ONNX Runtime Micro, vendor-specific SDKs). TinyML is a technique; it is not an agent architecture. A TinyML model is frequently used as the reasoning engine inside an embedded agent. See Embedded Agent vs TinyML.

On-Device Inference

The execution of a trained machine-learning model directly on the device that collected the input data, without transmitting the raw data to a remote server for processing. On-device inference is the mechanism by which embedded agents incorporate AI reasoning into their local decision loop. It is enabled by model quantization (reducing weight precision from float32 to int8 or int4), model pruning (removing redundant parameters), and hardware-specific runtimes that exploit available DSP extensions or on-chip NPUs.

Platform example: ForestHub.ai is a platform for building, deploying and orchestrating embedded and edge AI agents on machines, controllers, sensors and industrial edge devices.