Agentic AI, defined plainly.

An agent trace is the step-by-step record of an AI agent's run: its inputs, retrieved context, model calls, tool calls, checks, approvals and outcomes, with timing. Traces make agent behaviour reviewable after the fact and are the basis for debugging, evaluation and audit.

Agentic AI is software that pursues a goal by planning steps, choosing tools and taking actions, rather than only generating a reply. An agentic system decides what to do next, observes the result and continues. Because it acts, it needs verification, permissions and a record of what it did.

An agentic operating system is the layer that runs AI agents the way an operating system runs programs: it schedules their work, manages state and memory, controls which tools each may use, verifies consequential actions and records everything. RUUT BRAIN and FORGE together form this layer in the RUUT system.

An AI agent is a program that uses a model to decide on and carry out actions toward a goal — calling tools, reading data, writing records — within the permissions it has been given. Agents differ from chatbots in that their output is an action in a system, not only text.

An AI verification layer checks AI decisions before they take effect: it screens inputs, enforces policy on proposed actions, validates outputs against evidence and rules, and routes decisions that need a person to human approval. In the RUUT system this layer is RUUT FORGE.

Edge AI runs models close to where data is produced — beside cameras, sensors or machines — instead of sending raw data to a distant data centre. It is used where latency, bandwidth, privacy or connectivity make central processing impractical, such as plant floors and remote sites.

Grounding means constraining an AI system's output to information from an identified, trusted source — documents, records or sensor data — instead of the model's general knowledge. A grounded answer can be checked against its source; an ungrounded one can only be trusted or doubted.

Human-in-the-loop describes an AI system in which people review or approve certain decisions before they take effect. Well-designed systems decide in advance which actions need approval, route them to the right person, and record who approved what, instead of asking a human to check everything.

Least-privilege tool access gives each AI agent only the tools and permissions its task requires, through an explicit allow-list. It limits the damage a wrong decision or a manipulated input can cause, because an agent cannot call systems or actions it was never granted.

The Model Context Protocol is an open protocol for connecting AI applications to external tools and data sources through a standard interface. A system exposes tools over MCP once, and any compatible AI client can use them, subject to the permissions the server enforces.

A model-agnostic system can use different AI models interchangeably — open-weight or commercial, small or large — without being rebuilt. It lets each task use the most suitable model and avoids lock-in to a single provider. RUUT BRAIN routes work across models in this way.

An on-prem LLM is a large language model hosted on infrastructure an organisation controls, typically an open-weight model. It lets sensitive text and documents be processed without leaving that environment. Its capability, hardware cost and maintenance must be weighed against hosted alternatives for each task.

On-premise AI runs on computing infrastructure inside an organisation's own facilities rather than in a third-party cloud. It keeps data within a controlled environment and can reduce dependence on external connectivity, at the cost of owning capacity planning, hardware and operations.

PPE detection uses computer vision to check whether people in a defined area are wearing required personal protective equipment, such as helmets or high-visibility vests. Results depend on camera placement, lighting and rules per zone, so deployments are configured and validated site by site.

Prompt injection is an attack in which text supplied to an AI system — by a user or hidden in retrieved content — tries to override the system's instructions. Defences combine input screening, least-privilege tool access, output checks and approval for consequential actions, because no single filter is complete.

Retrieval-augmented generation is a technique in which a system first retrieves relevant passages from a trusted source and then has a model answer using those passages. It grounds answers in specific documents and lets each claim be traced to a source, reducing unsupported statements.

The RUUT Agent Loop is the four-stage path every RUUT agent follows: Perceive (VISION, SENSE), Reason (RUUT BRAIN), Verify (FORGE) and Act (agents such as FOREMAN). Using one loop for every agent makes behaviour predictable and makes every step traceable, from signal to action.

Sovereign AI is AI whose operating organisation controls where it runs, which models it uses, what data it can access and who approves its consequential actions. It usually implies on-premise or private-cloud deployment, model choice and enforced governance rather than dependence on one external provider.