Glossary, updated November 2026

The agentic AI glossary.

The words that come up in every conversation about AI agents, defined briefly and without the sales pitch. Sources for the standards are in the field notes on the agentic web.

AI agent: Software that takes a goal, chooses the tools it needs and acts until the goal is met, asking a person when a step needs approval. Read the guide.
Agentic AI: AI systems that plan and take actions toward a goal, as opposed to systems that only produce an answer for a person to act on.
Agentic web: The part of the internet where software agents read pages, compare options and complete tasks on a person’s behalf, using open standards built for them. Read the field notes.
Model Context Protocol (MCP): An open protocol, introduced by Anthropic in November 2024, that gives AI applications one standard way to connect to tools and data.
MCP server: A small service that exposes a system, such as a database, a booking tool or a document store, to agents through the Model Context Protocol.
Agent2Agent (A2A): An open protocol, started by Google in 2025 and now hosted by the Linux Foundation, that lets agents from different vendors find each other and hand over tasks.
Agent card: A machine-readable file, published at a well-known address on a site, that says what an agent can do and how to reach it.
llms.txt: A proposed plain-text file at the root of a website that gives language models a short, curated guide to the pages that matter most.
WebMCP: A proposed browser standard that lets a web page offer its own functions as tools to an agent working in the browser, so the agent does not have to guess at buttons and forms.
Agent Skills: Folders of instructions and scripts that an agent loads only when a task calls for them. Published by Anthropic as an open format in 2025.
Content Signals: An addition to robots.txt, proposed by Cloudflare in 2025, that lets a site state how its content may be used: for search, as input to AI answers, or for training.
Web Bot Auth: A proposed standard for bots and agents to sign their requests, so a website can verify who is calling before it answers.
Context engineering: The practice of deciding what an agent reads before it answers: which sources, in what order and with what evidence attached.
Context layer: The system that sits between an organisation’s information and its agents and decides what each agent is shown. Spiderbrain is ours. See Spiderbrain.
Retrieval-augmented generation (RAG): A method in which relevant passages are fetched from a store of documents and given to a language model along with the question, so the answer rests on them.
Deterministic retrieval: Retrieval that returns the same ranked context for the same question over the same data, every time. It makes an agent’s reading reviewable.
Grounding: Tying what an agent says to a specific source it was given, so a person can check the claim against the original.
Hallucination: A confident statement from a language model that is not supported by its sources or by fact. Grounding and review are the working defences.
Tool use: A model’s ability to call a function, such as a search, a lookup or a booking, and use the result. Also called function calling.
Autonomy level: How far an agent may go on a given task without a person: observe, draft, act with approval, or act within limits. How we set it.
Human in the loop: A design in which a named person reviews or approves an agent’s step before it takes effect.
Audit trail: The record of what an agent read, what it did and who approved it, kept so the work can be reviewed later.
Evaluation (evals): A fixed set of test cases an agent is run against before and after every change, to show whether it got better or worse.
Prompt injection: An attack in which instructions are hidden in content an agent reads, such as a web page or an email, to make it act against its owner’s wishes.
Multi-agent system: Several agents with separate jobs that pass work to one another, each with its own tools and limits.
Answer engine optimisation: Shaping a site’s content and structure so AI answer engines can find it, quote it accurately and cite it. Also called generative engine optimisation.

Start here

Bring us the operation that slows you down.

We will map it with you and say plainly where agents help, where they do not, and what it takes to run them safely.