AI agent terms, defined plainly.

What each term means, why it matters in regulated industries, and how it shows up in the Dynamiq platform.

Agent memory: The system that lets an AI agent recall earlier conversation turns or durable facts across separate runs, scoped to a user and session.
Agent orchestration: The layer that decides which agent or step runs next, in what order and with what shared state, when a task needs more than one agent.
Agent sandbox: An isolated, disposable virtual machine an agent can use to run shell commands, execute code and write files, separate from production systems.
Agentic AI: AI systems that pursue a goal by planning steps, calling tools and acting on the results, rather than only generating a response to a single prompt.
Agentic memory: Memory an AI agent manages itself, deciding which facts to save, update or recall across sessions, instead of only replaying past conversation.
Agentic workflow: A business process modeled as a graph of steps where at least one step is an agent that reasons and chooses its own actions, not a fixed script.
AI agent: A system that pairs a large language model with tools, memory and a control loop so it can plan and carry out multi-step tasks toward a goal.
AI agent platform: Software for building, deploying, governing and monitoring AI agents in one place, from models and tools to approvals, evaluations and traces.
AI gateway: A single endpoint that routes chat completion requests to many model providers behind one credential, so applications do not hold provider keys.
AI governance: The policies, controls and evidence an organization uses to decide how AI may be used, who is accountable, and how its behavior is checked over time.
AI guardrails: Checks placed around a model or agent, such as input screening and output validation, that flag or block unsafe behavior before it reaches anyone.
AI observability: The practice of recording every step an agent or workflow takes so its behavior, cost and failures can be inspected after the fact.
Air-gapped deployment: A deployment with no network path to the public internet, used when data must never leave a physically or logically isolated environment.
Approval gate: A control on a specific workflow step that pauses execution until a person approves, edits or rejects it, every time that step runs.

Barge-in and turn detection: How a voice agent knows a caller has stopped talking, and how it stops speaking immediately if the caller interrupts.

Chunking: Splitting a document into smaller pieces before it is embedded and indexed, so retrieval returns focused passages instead of whole files.
Computer use agent: An agent that operates a computer the way a person would, controlling a browser or desktop through clicks, typing and screenshots.
Context window: The maximum amount of text, measured in tokens, a model can read and generate in one call, including the prompt, history and output.
Conversational AI: Technology that lets people talk with software in natural language, by text or voice, and have it understand the request and respond or act on it.

Data residency: A guarantee about which country or region a system stores and processes its data in, used to meet legal or contractual requirements.
Document parsing: Converting a file such as a PDF, scan or slide deck into clean, structured text, typically Markdown, that a model or index can use.

Embeddings: Numeric vectors that represent the meaning of a piece of text, so pieces with similar meaning end up close together in vector space.
EU AI Act: The European Union's risk-based regulation of AI systems, Regulation (EU) 2024/1689, with obligations that scale to how risky a given use is.

Fine-tuning: Further training a pretrained model on a smaller, specific dataset so it performs better on a narrower task, tone or domain.
Forward-deployed engineer: An engineer who works inside a customer's own environment to build and ship the customer's first production use of a platform.

GDPR: The EU's General Data Protection Regulation, Regulation (EU) 2016/679, governing how personal data of people in the EU and EEA is handled.
GraphRAG: A retrieval method that answers questions by walking a knowledge graph of entities and relationships, not only by matching text similarity.
Grounding: Tying a model's answer to specific, retrievable source material, so the answer can be traced back to a document rather than training data.

Hallucination: An answer a model states with confidence that is not supported by its source material or by fact, presented as if it were true.
HIPAA: The US Health Insurance Portability and Accountability Act of 1996, whose Privacy and Security Rules govern protected health information.
Human in the loop: A design pattern where a person reviews, confirms or supplies information at a defined point in an otherwise automated process.

Intelligent document processing: Using AI to read unstructured documents such as scans, PDFs and forms and turn them into structured, validated data fields.

Knowledge graph: A database of entities and the relationships between them, such as a person working at an organization, used to answer connection questions.
KYC and AML automation: Using software, including AI, to run the identity checks and transaction monitoring that anti-money-laundering and know-your-customer rules require.

LLM as a judge: Using a large language model to score another model's output against a written rubric, in place of, or alongside, a human reviewer.
LLM evaluation: Scoring an LLM or agent's outputs against test data with defined metrics, so changes in quality are measured rather than judged by feel.
LLM inference: Running a trained model to produce an output for a given input, as opposed to training the model in the first place.
LLM tracing: A recorded, node-by-node execution tree of one run, including every prompt, tool call, retrieved document and token count.
LoRA: Low-Rank Adaptation, a fine-tuning method that trains a small set of extra weights on top of a frozen base model, not the whole model.

Model Context Protocol (MCP): An open protocol that lets an AI agent discover and call tools exposed by an external server over a standard interface.
Model routing: Automatically or programmatically directing a request to one of several possible models, based on name, cost, capability or availability.
Multi-agent system: Two or more agents working together on one task, each with a narrower role, coordinated by delegation or an explicit control flow.

Online evaluation: Automatically scoring a sampled share of a deployed system's live production traffic on an ongoing basis, not a one-time test.
Open-weight model: A model whose trained weights are published for anyone to download and run, unlike a model only reachable through a hosted API.

PII detection: Automatically flagging personally identifiable information, such as names, emails or account numbers, before it reaches a model.
Prompt injection: An attempt, hidden in user input or a retrieved document, to override a model's instructions and make it act against its role.

Reranking: Re-scoring and reordering a first pass of retrieved documents with a more precise model, so the most relevant ones reach the model first.
Retrieval-augmented generation (RAG): Retrieving relevant passages from your own documents and giving them to a model as context, so its answer is grounded in that material.
Role-based access control (RBAC): Granting permissions based on a person's assigned role, such as admin or member, rather than configuring access for each person.

Self-hosted AI platform: An AI platform installed and run inside an organization's own infrastructure, cloud account or data center, not a vendor's shared cloud.
Single sign-on (SSO): Letting people log in to a platform using the identity and credentials their organization already manages, not a separate password.
SIP trunking: A connection between a phone network and a voice application over the internet using the Session Initiation Protocol, not analog lines.
SOC 2: An audit standard from the AICPA evaluating a service organization's controls against the Trust Services Criteria: security, plus optional categories.
Sovereign AI: AI that an organization or country controls end to end: where it runs, which models it uses, where its data and records live, and who operates it.
Speech to text: Converting spoken audio into written text in real time, the first stage of a pipeline voice agent's turn.
Structured output: Constraining a model's final answer to a JSON schema, so downstream code receives typed, parseable data instead of free-form prose.

Text to speech: Converting written text into spoken audio, the final stage of a pipeline voice agent's turn, or one channel of a realtime model.
Tool calling: A model's ability to invoke a defined function or external action mid-response, then read the result back before continuing.

Vector database: A database built to store embeddings and search them by similarity, returning the nearest vectors to a query, not exact matches.
Voice agent: A real-time conversational agent that talks with people over audio, on a phone call, in a browser, or over a raw audio connection.

WebRTC: An open, browser-native protocol for real-time audio and video that lets a web page or app stream microphone audio to a voice agent.

See an agent on your own workflow.

Bring a process and its documents. Our engineers will show you how Dynamiq runs it, in your environment or ours.

Start free Talk to the team