The AI agent glossary.

A software system powered by an LLM that perceives its environment, plans actions, and executes them — usually across multiple steps and tools.

The core control flow of an agent: observe → reason → act → observe, repeated until the goal is met or a stop condition fires.

An agent that plans, executes, and finishes a multi-step task without asking for human approval between steps.

A publicly-shared, standardized eval suite used to compare models and agents across a uniform task — SWE-bench, MMLU, GAIA, etc.

An agent capability where the LLM drives a real web browser to read, click, and fill forms on live websites.

A deployment pattern where you supply your own model API key to the agent — token costs are billed to you directly, the agent vendor charges only for the software.

A prompting technique that asks the model to lay out its reasoning step-by-step before committing to an answer — improves accuracy on multi-step tasks.

An agent capability for writing and running code in a sandboxed environment — usually Python — to compute, transform data, or test hypotheses.

The maximum number of tokens a model can consider at once — covers the system prompt, conversation, tool results, and the answer being generated.

An AI tool that suggests changes inline and waits for the user to accept — the human stays in the driver's seat.

In support agents: the percentage of customer contacts the agent resolves fully without escalating to a human.

A systematic test that measures agent performance on a fixed set of inputs — the agent equivalent of a test suite.

A pricing model where the agent has a useful free tier with paid plans for higher usage, more features, or commercial use.

The current generation of state-of-the-art LLMs — typically the largest models from OpenAI, Anthropic, Google, and a small number of others.

When an LLM generates content that sounds plausible but is factually wrong or fabricated — a citation that doesn't exist, a function that isn't in the API.

A workflow pattern where the agent pauses for human approval at one or more checkpoints before continuing.

Model Context Protocol — an open standard that lets any agent talk to any tool or data source through a single protocol.

The mechanism by which an agent remembers information across sessions — usually a vector store or structured key-value cache.

An architecture where several specialized agents collaborate on the same task — each handles a sub-goal and they coordinate through a shared workspace.

A deployment where the agent runs entirely on infrastructure the customer controls — no agent code or customer data leaves the customer's network.

An agent whose source code is publicly licensed (MIT, Apache, AGPL) — you can self-host, fork, and audit.

A pricing model where you pay per completed task — per PR generated, per ticket resolved, per email drafted — rather than per seat or per month.

The phase where an agent decomposes a goal into a structured sequence of sub-tasks before executing any of them.

Retrieval-augmented generation — pulling relevant documents from a knowledge base before generating, so the LLM grounds its answer in your data.

A class of LLM (o3, Claude Sonnet 4.6, Gemini 2.5 reasoning) that produces a long internal chain of thought before responding — trading latency for accuracy on hard problems.

An agent that plans and executes most steps unsupervised but pauses for approval before anything irreversible.

A model feature that constrains output to a specific JSON schema, making LLM responses safely parseable by downstream code.

A flat-rate monthly price per user — the dominant pricing model for agents aimed at individual contributors.

The initial instruction text given to an LLM that sets its persona, tools, constraints, and default behavior for the session.

Total cost of ownership — the all-in cost of running an agent including subscription, token spend, ops time, and integration work.

The ability of an LLM to invoke external functions — APIs, shell commands, internal services — instead of just generating text.

An agent capability for understanding images, screenshots, and video — letting the model reason over visual content.

An agent capability for taking phone calls, holding spoken conversations, and triggering actions from voice input.