The vocabulary of
AI agents.

An AI agent runtime is the long-running software that hosts an AI agent — it owns the model connection, the tools the agent can call, its memory, and the loop that turns a request into actions.

Unlike a chat window, it keeps state across conversations and runs work in the background — the foundation a professional agent like lynox needs to operate a business.

A professional agent is an AI agent built to run a business’s recurring operational work — it remembers clients and context across conversations and executes scheduled tasks unattended.

lynox is a professional agent — as opposed to a personal assistant, which optimizes for one-off chats rather than persistent business operations.

An AI personal assistant is a chat-first agent optimized for individual, ad-hoc tasks; it typically doesn’t persist structured business context or run unattended work.

Most consumer AI tools are personal assistants — great for individual productivity, not built to operate a business.

A knowledge graph is a typed store of entities (people, companies, deals, projects) and the relationships between them, giving an AI agent structured, persistent context instead of a flat chat history.

It lets the agent answer "what’s open with Acme?" by reading named facts, not by re-reading every past message.

The agent loop is the cycle in which a model reads context, calls a tool, reads the result, and repeats until the task is done.

It’s what separates an agent from a single prompt-and-response: the model acts, observes, and adjusts over multiple steps.

Self-hosted AI means running the AI application on infrastructure you control — your own server or cloud account — rather than a vendor’s multi-tenant cloud, so your data stays in your environment.

The AI model provider still sees the conversation context, but on self-host no application vendor sits in the middle storing your data.

BYOK means you supply your own API key for the AI model provider, so you’re billed by the provider directly instead of through the application vendor.

It keeps your model usage transparent and uncapped by the app, and your key under your control.

Workflow capture is turning a multi-step task an agent just performed into a reusable, parameterized workflow that can be re-run or scheduled.

Instead of re-describing a recurring job each time, you save it once and the agent repeats it on demand or on a schedule.

A background worker is the part of an agent runtime that executes scheduled or triggered tasks independently of any open chat session.

It’s what lets an agent run a Monday-morning report on a cron schedule whether or not you’re online.

EU data residency means personal and business data is processed and stored within the European Union — a common requirement for GDPR compliance when choosing AI providers.

Choosing an EU-based model provider (e.g. Mistral in Paris) avoids a transatlantic data transfer and reduces US CLOUD Act exposure.

Data sovereignty is the principle that data is subject to the laws of the country where it’s stored, and that its owner keeps control over where it lives and who can access it.

Self-hosting plus an EU-based AI provider is a practical way for a business to keep data sovereignty.

Provider-agnostic means the software can run on multiple AI model providers — e.g. Anthropic, Mistral, or any OpenAI-compatible endpoint — and switch between them without re-architecting.

It protects you from vendor lock-in: if a provider changes pricing or policy, you switch the key, not the system.

A model tier is a capability-and-cost band an agent routes work to — for example a deep tier for hard reasoning and a fast tier for cheap routine steps — independent of which vendor supplies the model.

Naming tiers by function (deep / balanced / fast) instead of a vendor’s brand keeps the setup provider-agnostic.

Prompt caching reuses the model’s cached internal state for an unchanged prompt prefix — the system prompt, tool schemas, and earlier history — across turns, cutting the cost and latency of long agent loops.

Because an agent re-sends its whole context every turn, caching can swing the real cost of running a loop by several times.

Tool use is an AI model’s ability to call external functions or APIs — search, send email, query a database — and use the results, rather than only generating text.

It’s the capability that makes an agent able to act in the world; the agent loop is built around repeated tool use.