Clear, source independent definitions of the terms behind edge AI, agents and industrial orchestration. Each entry answers one question directly, with a short ForestHub note where it helps.
Foundations
The core ideas behind running artificial intelligence on local edge devices instead of the cloud.
Edge AI
Edge AI is the practice of running artificial intelligence directly on local devices such as sensors, gateways, machines or embedded boards, instead of sending data to a remote cloud service. The model runs where the data is produced, so decisions happen with low latency and without a network round trip. Edge AI is used when bandwidth, privacy, autonomy or response time make a cloud dependency impractical.
At ForestHub
ForestHub is an orchestration platform for edge AI. The engine runs as a Docker image on Linux edge devices and coordinates models, tools and industrial signals locally.
An edge agent is an AI agent that runs on a local edge device rather than in the cloud. It perceives signals from its environment such as sensors or industrial protocols, reasons over them with a model, and acts on the physical system without a round trip to a datacenter. Edge agents are useful where autonomy, latency and data locality matter.
At ForestHub
At ForestHub the workflow is a deterministic graph and the model is one node among many, so an edge agent stays inspectable, replayable and bounded.
An embedded agent is an AI agent that operates inside a resource constrained embedded system, working within tight limits on memory, compute and power. It combines local perception and decision making on hardware that often has no reliable cloud connection. Embedded agents bring autonomous behavior to machines, appliances and field equipment.
At ForestHub
ForestHub targets Linux based edge hardware, so embedded agents can run the full orchestration engine rather than a stripped down runtime.
On-device inference is the execution of a trained machine learning model on the same device that captures the input, producing the prediction locally. No raw data leaves the device for a prediction to be made, which protects privacy and removes network latency. It is the runtime counterpart of training, which typically runs separately on more powerful hardware in a data center.
At ForestHub
ForestHub routes inference per node to local or remote models, so a workflow can keep sensitive steps on the device.
Edge AI and cloud AI describe where an AI model runs. Cloud AI executes models in a remote datacenter with large compute and centralized data, while edge AI executes them on local devices near the data source. The trade off is capacity and convenience in the cloud against latency, privacy and autonomy at the edge.
At ForestHub
ForestHub lets a single workflow mix both, calling a large cloud model for hard cases and a local model for routine ones.
Inference is the phase in which a trained model is applied to new input to produce an output such as a classification, a prediction or generated text. It is distinct from training, the earlier phase where the model learns its parameters. Inference is what runs in production every time a model answers a request.
At ForestHub
In a ForestHub workflow each model call is an inference step inside the graph, with its input and output recorded.
Quantization is a model compression technique that lowers the numerical precision of the model weights, for example from 32 bit floating point to 8 bit integers. The smaller representation cuts memory use and speeds up computation, which makes large models practical on constrained edge hardware. The cost is a usually small loss of accuracy.
TinyML is the field of running machine learning models on very low power microcontrollers and embedded devices with only kilobytes of memory. It focuses on extreme efficiency so that inference can happen on battery powered or always on hardware. Typical uses include keyword spotting, anomaly detection and simple sensor classification.
Language models and the retrieval techniques that ground them, from small on device models to large cloud models.
Small Language Model (SLM)
A small language model is a language model with a relatively small number of parameters, designed to run efficiently on limited hardware including edge devices. It trades some general capability for lower memory use, faster responses and the ability to run locally or offline. Small language models are often specialized or fine tuned for a narrow task.
At ForestHub
ForestHub can route routine workflow steps to a small local model and escalate only hard cases to a larger one.
A large language model is a neural network trained on very large text corpora to predict and generate language, with billions of parameters. It can follow instructions, answer questions, summarize, classify and write code across many domains. Large language models usually run in the cloud because of their compute and memory needs.
At ForestHub
In a ForestHub graph an LLM is one node among many, called only where it adds value and always inside a bounded, auditable step.
Retrieval-augmented generation is a technique that gives a language model relevant external documents at query time instead of relying only on what it learned during training. A retrieval step finds matching passages from a knowledge base, and the model generates its answer grounded in that context. RAG improves accuracy and lets a model use private or up to date information.
At ForestHub
ForestHub workflows can include a retrieval node backed by a local knowledge base, so grounding data never has to leave the edge device.
Small and large language models differ mainly in size, capability and where they can run. A large model is more capable and general but needs cloud scale compute, while a small model is narrower but runs locally with low latency and cost. Many systems combine both, using the small model by default and the large model only when needed.
At ForestHub
ForestHub makes this split explicit as routing in the graph, often with a confidence routed cascade.
A context window is the maximum amount of text, measured in tokens, that a language model can consider at once when producing an output. Everything the model reasons over, including the prompt and any retrieved documents, must fit inside this window. A larger context window allows longer inputs but costs more compute and memory.
Fine-tuning is the process of further training a pretrained model on a smaller, task specific dataset so it performs better on that task or domain. It adapts general knowledge to specialized vocabulary, formats or behavior without training a model from scratch. Fine-tuning is one way to specialize a small model for an edge use case.
A token is the basic unit of text that a language model processes, often a word, a part of a word or a punctuation mark. Models read and generate text as sequences of tokens, and limits such as the context window and pricing are usually counted in tokens. Roughly, one token corresponds to about four characters of English text.
An embedding is a numerical vector that represents the meaning of a piece of text, an image or other data in a way a computer can compare. Items with similar meaning end up close together in the vector space, which makes search and retrieval by similarity possible. Embeddings are the foundation of vector search and retrieval augmented generation.
How AI agents are built and orchestrated, from a single agent to graph based multi step workflows.
AI Agent
An AI agent is a software system that uses a model to decide and take actions toward a goal, rather than only answering a single question. It can observe a situation, choose among tools or steps, act, and react to the result over multiple turns. Agents range from simple tool callers to systems that plan and adapt.
At ForestHub
At ForestHub an agent is expressed as a graph of nodes, so its decisions and actions are visible and bounded rather than hidden inside a loop.
An agentic workflow is a defined sequence of steps in which one or more AI agents carry out a larger task, often mixing model calls, tools, data sources and human checks. Unlike a single prompt, it coordinates multiple actions toward an outcome and can branch based on intermediate results. Agentic workflows make complex automation repeatable and inspectable.
At ForestHub
ForestHub authors agentic workflows as graphs in a visual builder and runs them with a deterministic engine on the edge.
Graph orchestration is an approach to building AI systems where the workflow is defined as an explicit graph of nodes and connections, and an engine executes that graph. Each node is a discrete step such as a model call, a tool, a condition or an actuator, and the edges define the flow of data and control. The graph makes the program structure visible and testable instead of emergent.
At ForestHub
This is the core of ForestHub. The graph is the program and the model is one node among many, which keeps execution deterministic and auditable.
The Model Context Protocol is an open standard that lets AI applications connect to external tools and data sources through a common interface. Instead of custom integrations per model, a server exposes capabilities that any MCP aware client can use. MCP standardizes how an agent discovers and calls tools.
A workflow node is a single step in an orchestrated workflow graph, representing one discrete operation such as a model call, a data query, a condition, a transformation or an action on a device. Nodes are connected by edges that pass data and control from one step to the next. Building a workflow means composing nodes rather than writing imperative code.
At ForestHub
In the ForestHub visual builder a workflow is assembled from typed nodes on a canvas.
Tool use, also called function calling, is the ability of a language model to invoke external functions or services instead of only producing text. The model decides when a tool is needed, supplies the arguments, and uses the returned result to continue. Tool use is what lets an agent query data, call an API or act on a system.
Agent to agent communication is the exchange of messages, tasks or results directly between autonomous AI agents so they can collaborate on a larger problem. Each agent can specialize, and a coordinating structure routes work and combines outcomes. A2A patterns underpin multi agent systems.
A multi agent system is a system in which several AI agents, each with its own role or specialty, work together to solve a task that is hard for a single agent. Agents may divide the work, review each other or operate in parallel, coordinated by a defined structure. The design adds capability but also raises questions of control and predictability.
At ForestHub
ForestHub expresses coordination as an explicit graph rather than an open ended conversation, which keeps a multi agent system bounded.
Human in the loop is a design in which a person reviews, approves or corrects the decisions of an AI system at defined points before they take effect. It keeps human judgment in control of high stakes or uncertain actions while still automating the routine work. The pattern is common where errors are costly or regulated.
At ForestHub
A ForestHub workflow can place an approval node before any actuating step, so a human gate is part of the graph.
Graph first and loop first are two ways to structure an AI agent. In a loop first design the model runs in an open loop and decides each next step at runtime, so the loop itself is the program. In a graph first design the control flow is an explicit graph authored in advance, and the model fills specific nodes, which makes the path auditable and bounded.
At ForestHub
ForestHub is graph first by design, which is why every decision and action in a workflow can be inspected and replayed.
Patterns and concepts that make AI agents predictable, inspectable and safe enough for physical systems.
Verification-Gated Actuation
Verification-gated actuation is a safety pattern in which an AI agent may propose an action but a separate, deterministic check must pass before that action reaches the physical system. The model decides, a rule based verifier validates, and only verified actions are executed while the rest are logged or escalated. It keeps a probabilistic model from directly driving a real world actuator.
At ForestHub
This is one of the ForestHub agent patterns, expressed as a verifier node between the model and any actuator.
A bounded agent is an AI agent whose possible actions are constrained in advance to a known, limited set rather than left open ended. The bounds define what the agent is allowed to do, so its behavior stays predictable even when the underlying model is not. Bounding is a core technique for using agents safely in industrial and physical systems.
At ForestHub
ForestHub bounds agents by construction, because the graph defines exactly which nodes and actions exist.
An audit trail is a complete, time ordered record of what a system did and why, kept so that actions can be reviewed, explained or replayed after the fact. For an AI agent it captures each input, decision, model output and resulting action. An audit trail is essential where accountability, debugging or regulation require knowing how a result was reached.
At ForestHub
Because a ForestHub workflow is a deterministic graph, each run records the path taken and the data at every node.
Shadow mode deployment runs a new AI agent alongside the existing system without letting it act, so its decisions can be compared against the trusted path. The agent sees real inputs and produces real outputs, but those outputs are logged instead of executed. After enough comparison the agent can be promoted with evidence that it behaves correctly.
At ForestHub
Shadow mode is one of the ForestHub patterns, used to validate an agent on live data before it actuates.
A confidence routed cascade is a pattern that sends easy cases to a small, cheap model and escalates only low confidence cases to a larger model or a human. The first stage classifies and, when sure enough, acts, while uncertain cases move up the cascade for more capable handling. It controls cost and latency without giving up accuracy on hard cases.
At ForestHub
ForestHub implements this as routing in the graph, often with a small local model as the first stage.
Replay testing is a method that records the real inputs and outputs of a production AI workflow and re runs them later in a test environment to check for regressions. Recorded model outputs can be replayed deterministically so downstream logic is tested without calling the model again. It makes non deterministic systems testable in continuous integration.
At ForestHub
ForestHub records workflow runs so they can be replayed, which is one of its core patterns.
AI as classifier is a conservative pattern in which the model only categorizes a situation into a fixed set of labels, and deterministic code decides what to do with each label. The model never acts directly, it only informs a rule based branch. Restricting the model to classification keeps its influence predictable and easy to validate.
At ForestHub
ForestHub uses this pattern to let a model label a situation while the graph decides the response.
A state machine augmented agent combines a model with an explicit state machine that defines which states exist and which transitions are allowed. The model can drive transitions, but only the ones the state machine permits, so the agent cannot enter undefined or unsafe states. It blends adaptive decision making with a predictable control structure.
At ForestHub
ForestHub graphs can encode states and allowed transitions, keeping an agent inside a defined lifecycle.
Guardrails are constraints placed around an AI system that prevent it from producing unsafe, invalid or out of scope outputs and actions. They can validate inputs and outputs, restrict tools, enforce formats or block disallowed behavior. Guardrails turn a flexible model into a component that behaves within known limits.
At ForestHub
At ForestHub, guardrails are structural, since the graph defines the only paths and actions available to an agent.
Deterministic execution means that, given the same inputs and state, a system always follows the same path and produces the same observable behavior. For AI systems it usually refers to the orchestration around the model being deterministic, even though the model itself is probabilistic. Determinism is what makes a workflow auditable, testable and safe to run on physical equipment.
At ForestHub
ForestHub runs workflows with a deterministic engine, so the model is the only probabilistic part and everything around it is reproducible.
The protocols, systems and use cases that connect edge AI to real industrial equipment.
MQTT
MQTT is a lightweight publish and subscribe messaging protocol designed for constrained devices and unreliable networks, widely used in the Internet of Things and industrial systems. Devices publish messages to named topics through a broker, and other devices subscribe to the topics they care about. Its small footprint makes it a common transport for sensor and telemetry data at the edge.
At ForestHub
The ForestHub engine speaks MQTT, so workflows can consume and publish messages on industrial message buses.
Modbus is a long established industrial communication protocol used to connect controllers, sensors and actuators, especially in manufacturing and building automation. It defines a simple request and response structure for reading and writing device registers over serial lines or TCP networks. Its age and simplicity make it one of the most widely supported protocols in the field.
At ForestHub
Modbus support is on the ForestHub roadmap. The engine speaks MQTT today.
OPC UA, short for OPC Unified Architecture, is a platform independent industrial standard for secure data exchange between machines, controllers and software systems. It provides a structured information model and built in security, which makes it a backbone of modern industrial interoperability. OPC UA is common in factories that need vendor neutral machine to machine communication.
At ForestHub
OPC UA support is on the ForestHub roadmap. The engine speaks MQTT today.
The Industrial Internet of Things is the application of connected sensors, machines and analytics to industrial operations such as manufacturing, energy and logistics. It links physical equipment to software so that data can drive monitoring, optimization and automation. IIoT differs from consumer IoT in its demands for reliability, safety and integration with existing industrial systems.
At ForestHub
ForestHub adds local AI orchestration to IIoT setups, so decisions can be made on the edge device next to the machine.
A programmable logic controller is a ruggedized industrial computer that controls machines and processes by executing control logic in real time. It reads inputs from sensors, runs a deterministic control program, and drives outputs to actuators, all built to survive harsh factory conditions. PLCs are the backbone of industrial automation.
At ForestHub
ForestHub runs alongside existing controllers and does not replace the deterministic control a PLC provides.
Predictive maintenance is a strategy that uses sensor data and analytics to predict when equipment is likely to fail, so maintenance happens just before a problem occurs. It aims to reduce both unplanned downtime and the waste of fixed schedule servicing. Models detect patterns such as vibration, temperature or current that precede faults.
At ForestHub
Predictive maintenance is a common ForestHub edge use case, with the analysis running locally on the machine data.
SCADA, short for supervisory control and data acquisition, is a category of industrial control system used to monitor and control large, distributed processes such as utilities, pipelines and factory lines. It gathers data from remote equipment, presents it to operators, and lets them issue control commands from a central place. SCADA systems emphasize visibility and supervisory control across many sites.
A digital twin is a virtual model of a physical asset, process or system that is kept in sync with its real counterpart using live data. It lets teams simulate, monitor and analyze the real thing without touching it, for example to test changes or predict behavior. Digital twins are used across manufacturing, energy and infrastructure.
