The AI agent glossary

Software that completes a task end to end, deciding the steps and using tools to act.

The model that gives an agent its language understanding and reasoning ability.

Short for large language model, the reasoning core of an agent.

Retrieval Augmented Generation: answering from your data, not just the model’s training.

Looking up relevant facts from your content before answering, to stay grounded.

A numeric representation of text that lets an agent measure how related two things are.

Storage that finds the most relevant chunks of your knowledge quickly.

The instructions that tell an agent how to behave and respond.

The craft of writing instructions that get reliable results from a model.

The always-on instructions that set an agent’s role, rules and tone.

Training a model further on your data to specialize it for a task.

How an agent triggers real actions in your tools by calling defined functions.

An agent using external systems to get work done, not just chatting.

Model Context Protocol: a standard way to give agents context and tools.

An open standard for connecting agents to data sources and tools.

A service that exposes tools and data to agents over the Model Context Protocol.

Several agents coordinating, each specialized, to handle a larger job.

Coordinating multiple agents or steps into one reliable workflow.

What an agent remembers across a conversation or over time.

The context an agent holds within a single conversation.

Knowledge an agent retains and recalls across sessions.

How much text a model can consider at once when responding.

The unit of text models process; usage and limits are measured in tokens.

A setting that controls how varied or predictable a model’s output is.

When a model states something false with confidence; RAG and guardrails reduce it.

Rules that keep an agent safe, on-topic and honest.

Tying an agent’s answers to real source data so they can be verified.

The model’s ability to break a problem into steps and work through it.

How an agent decides the sequence of actions to reach a goal.

A prompting approach where the model reasons step by step before answering.

A pattern where an agent alternates reasoning and taking actions.

An agent that can pursue a goal with minimal step-by-step human input.

Keeping a person in control of key decisions an agent makes.

Turning a caller’s voice into words the agent can read.

Speech to text, the input side of a voice agent.

Turning the agent’s reply into natural spoken audio.

Text to speech, the output side of a voice agent.

Agents that hold natural spoken conversations over the phone or app.

The field of getting computers to understand human language.

Short for natural language processing.

Working out what a user actually means, beyond the literal words.

What the user is trying to accomplish in a message or call.

A specific detail an agent extracts, like a date, name or amount.

Optical character recognition: reading text from scanned documents and images.

Reading, understanding and extracting data from real-world documents.

Getting models to interpret images and video.

The collection of documents an agent answers from.

A connected map of facts and relationships an agent can reason over.

Splitting documents into pieces so they can be embedded and retrieved.

Finding results by meaning rather than exact keyword match.

Reordering retrieved results so the most relevant come first.

A list of numbers representing the meaning of a piece of content.

How quickly an agent responds; lower is better for voice and chat.

How much work an agent can handle in a given time.

Running a model to produce an output from an input.

The trained system that produces predictions or text.

A large, general model that many applications build on top of.

One of the most capable current models available.

A model whose weights are publicly available to run and adapt.

A model that works with more than one type of input, like text and images.

A process where agents take actions across tools to reach an outcome.

Automating a multi-step process so it runs without manual work.

The event that starts an agent or workflow.

A message one system sends another when something happens.

A defined way for software systems to talk to each other.

A common style of web API using standard HTTP methods.

A query language that lets clients request exactly the data they need.

A connection that lets an agent work inside one of your tools.

Customer relationship management software where sales and contact data lives.

Software teams use to manage customer support tickets.

A single customer request tracked in a help desk.

Resolving a request before it becomes a ticket for your team.

A potential customer who has shown some interest.

Deciding whether a lead is worth a rep’s time and routing it.

Ranking leads by how likely they are to convert.

The set of deals moving toward a close.

When a customer stops using or paying for a product.

Customer satisfaction, often measured by a short post-interaction survey.

Net promoter score, a measure of how likely customers are to recommend you.

Service level agreement: a promise about response time or uptime.

The share of time a service is available and working.

Putting an agent live so it starts doing real work.

Running software inside your own infrastructure rather than a vendor’s cloud.

A cloud environment dedicated to one organization.

Virtual private cloud: an isolated section of a cloud provider you control.

Control over which region your data is stored and processed in.

Encoding data so only authorized parties can read it.

Protecting stored data by keeping it encrypted on disk.

Protecting data while it moves between systems.

Single sign-on: logging in once to access many systems.

A standard that enables single sign-on between systems.

Role-based access control: permissions based on a user’s role.

A record of actions taken, used for review and compliance.

A security compliance standard covering how a company handles data.

European data-protection law giving people rights over their data.

US law governing the privacy and security of health information.

Security standards for handling payment card data.

Data processing agreement: a contract on how a vendor handles your data.

A third party a vendor uses to help deliver its service.

Personally identifiable information about an individual.

Protected health information covered by HIPAA.

Collecting and using only the data you actually need.

Building and using AI in ways that are safe, fair and accountable.

The policies and controls that keep AI use responsible and auditable.

Systematic unfairness in a model’s outputs.

How well you can understand why a model produced an output.

Testing an agent’s quality against known cases.

A standard test used to compare models or agents.

A model that checks another agent’s output for safety or policy.

A cap on how many requests can be made in a time window.

Handling many items together on a schedule rather than one at a time.

Returning an agent’s response progressively as it is produced.

Detecting the emotional tone of a message.

Condensing long content into the key points.

Sorting inputs into categories, like routing a ticket.

Pulling structured fields out of unstructured text.

Tailoring an agent’s responses to the individual user.

Handing a conversation to a human when the agent should not proceed.

Passing a conversation to a person with the full context.

The date after which a base model has no built-in knowledge.

An attack that tricks an agent into ignoring its instructions.

An isolated environment where an agent can act safely for testing.

What it costs to run one full agent interaction.

How quickly a deployed agent starts delivering results.