Words Atarla tries not to use loosely.

Core terms.

• Layer: the control software itself, the part of Atarla that holds authorized context, enforces policy, and coordinates agents, independent of where it runs.
• Node: supported hardware, planned rather than shipping today, meant to run the layer on a business's own premises.
• Cloud: Atarla operating the layer on a business's behalf, as one deployment option among several, not the definition of the company.
• Agent: a narrow, identified actor with declared tools and a limit. A model may recommend or plan; the agent is what actually acts, and only within what it is permitted.
• Authorized memory: a durable context record that carries an owner, a scope, a source, a sensitivity classification, an access policy, and a retention rule, so it can be governed, moved, or safely forgotten.
• Secrets boundary: the separation between credentials, API keys, and private keys and everything else. Secrets reach an agent or tool only for the minimum purpose and duration required, and are never stored as memory.
• Deterministic policy: permission logic enforced outside the model, so whether an action is allowed does not depend on what a model decided to output.
• Local-first: work starts from the business's own policy and the capability closest to it, and reaches an outside model only when there is a stated reason. It is a control order, not a promise that everything works offline.
• Bounded action: an action an agent may take only within an explicit scope of data, tools, and consequence, with the option to require human approval before it executes.
• Provenance: the recorded origin of a piece of context or an action, including which source produced it and which policy applied, so a decision can be traced back to what caused it.
• Deployment location: where the layer physically runs, whether that is Atarla Cloud, a business's own infrastructure, or an Atarla Node. Location is a deployment decision, not a different product.