Glossary

An AI model that plans and takes actions with tools, such as searching, editing files or calling an API, instead of only answering questions.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

An open standard for connecting AI agents to tools and data. An MCP server offers the tools; the agent calls them through an MCP client.

Explained in MCP specification changes, explained

Labels on an MCP tool that say whether it only reads data, may change or delete it, can safely be repeated, or talks to the outside world. They are hints for the client, not guarantees.

Explained in MCP tool design: fewer, safer, testable tools

The standard way for an application to get permission to act for a user without seeing their password. MCP uses it for sign-in on remote servers.

Explained in Delegated access for agents: OAuth 2.1, PKCE and scoped tokens

An extra check during OAuth sign-in that stops a stolen login code from being used by someone else. Mandatory for MCP clients.

Explained in Delegated access for agents: OAuth 2.1, PKCE and scoped tokens

One permission inside an access token, such as read invoices or send email. Narrow scopes keep an agent from doing more than its task needs.

Explained in Delegated access for agents: OAuth 2.1, PKCE and scoped tokens

An agent acting on behalf of a signed-in user, with its own identity and rights that can never exceed the user's, so logs show both who asked and what acted.

Explained in Delegated access for agents: OAuth 2.1, PKCE and scoped tokens

Forwarding a user's access token from one system to another. Forbidden in MCP: a server must use its own credentials for upstream systems.

Explained in Delegated access for agents: OAuth 2.1, PKCE and scoped tokens

Instructions hidden in content an agent reads, such as a web page, an issue or a support ticket, that make it do something the user did not ask for.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

Simon Willison's name for the risky combination of access to private data, exposure to untrusted content and a way to communicate externally. An agent with all three can be tricked into leaking data.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

OWASP's term for an AI system that can take damaging actions because it has too much functionality, too many permissions or too much autonomy.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

The share of tasks an agent solves on every one of k tries. It measures reliability, unlike a single successful run.

Explained in AI agent reliability benchmarks, tracked

METR's measure of the length of task, in a skilled person's working time, that a model completes with a given success rate, such as 50% or 80%.

Explained in AI agent reliability benchmarks, tracked

Using language models or coding agents to move a codebase to a new framework, library or language version, with automatic checks on every change.

Explained in AI code migrations: what Airbnb, Google and Amazon actually reported

Giving a language model relevant passages from your own documents at question time, so it answers from them instead of from memory alone.

Explained in GraphRAG vs RAG: what the evidence actually says

Finding passages whose meaning is close to the question, using numeric representations of text called embeddings. Good for concepts, weak for exact codes and identifiers.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

The classic keyword ranking method behind most search engines. Strong on exact words, product codes and acronyms.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

Combining keyword ranking and vector search, usually followed by a reranker. In Microsoft's tests it beat either method alone.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

A second model that reorders the top search results by how well they answer the question.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

Data stored as entities and the relations between them, such as part, product and customer, so connected facts can be followed step by step.

Explained in Five questions before you build a knowledge graph for RAG

RAG that builds a knowledge graph from your documents with a language model and retrieves through it. Helps most with questions that connect facts across documents.

Explained in GraphRAG vs RAG: what the evidence actually says

A question whose answer needs several connected facts, often from different documents, such as which customers are affected by a recalled part.

Explained in Five questions before you build a knowledge graph for RAG

In Microsoft GraphRAG, summaries written for clusters of related entities in the graph. Global search answers from them.

Explained in The real cost of GraphRAG: indexing, queries and keeping the graph fresh

GraphRAG's two query modes: global search answers whole-corpus questions from community summaries; local search starts from the entities in the question.

Explained in The real cost of GraphRAG: indexing, queries and keeping the graph fresh

A Microsoft variant that skips summarising the graph up front and does more work at query time, to cut indexing cost.

Explained in The real cost of GraphRAG: indexing, queries and keeping the graph fresh

A search quality score from 0 to 1, often shown as 0 to 100, that rewards putting the most relevant results first.

Explained in Agentic AI from demo to production: what the evidence says about the architecture

A study that assigns people or tasks at random to work with or without AI, so differences in results can be attributed to AI.

Explained in Does AI make developers faster? What the controlled studies say

A randomised trial run in normal work rather than in a lab, such as giving a random group of developers an AI tool for months.

Explained in Does AI make developers faster? What the controlled studies say

The range within which the true effect probably lies. A wide interval, or one that includes zero, means the result is uncertain.

Explained in Does AI make developers faster? What the controlled studies say

Making decisions from the best available evidence in six steps: ask, acquire, appraise, aggregate, apply and assess.

Explained in A GraphRAG business case, done the evidence-based way

A result agreed before a pilot starts that decides whether the project continues or stops.

Explained in A GraphRAG business case, done the evidence-based way

The European Union's law on artificial intelligence, Regulation (EU) 2024/1689, with obligations by risk level and by role, applying in stages from 2025 to 2028.

Explained in EU AI Act dates and guidance, tracked

Under the EU AI Act, AI used in sensitive areas such as employment, education or critical infrastructure, or in regulated products. Its obligations apply from December 2027 and August 2028.

Explained in EU AI Act dates and guidance, tracked

A European standard cited in the Official Journal; following it gives a presumption of conformity with the law. EN 18286 is the first one under the AI Act.

Explained in EU AI Act dates and guidance, tracked

The international standard for an AI management system: policies, roles, risk treatment and audits that an independent body can certify.

Explained in AI transformation frameworks compared: which one is for what

The US National Institute of Standards and Technology's voluntary framework for managing AI risk, built on four functions: govern, map, measure and manage.

Explained in AI transformation frameworks compared: which one is for what

A company's estimate of how many human agents' work an AI system takes on. It is an estimate of avoided work, not a count of jobs cut.

Explained in Klarna's AI customer service: the 700-agent claim and the U-turn

The share of tasks an agent must get right for its attempts plus a person checking them to cost no more than the person doing the work alone. Above it the agent saves money; below it, it adds cost.