Glossary
The terms we use on this site and in our work, each defined in one sentence that stands on its own, with the service and the notes where it comes up.
An AI agent is a language model that can call tools to act on real systems, such as searching, reading a record or placing an order, on behalf of a person.
Most of an agent's quality is in its tools, not the model: a few of them, named and described so the model picks the right one, each returning compact results. Anything irreversible should wait for a person's approval.
A multi-agent system is one in which several AI agents, each with its own role and tools, work on the same task.
Add an agent when the work has a genuinely different mandate, not because the task is big. DevGuardian AI runs three, an Architect, a QA reviewer and a Security reviewer, because each owns a different question about the same change.
The Model Context Protocol (MCP) is an open standard that lets AI hosts such as ChatGPT, Claude, VS Code and Gemini CLI connect an agent to a product's tools.
A product takes part through an MCP server. The specification changed often in its first two years; its 2026-07-28 revision is a stateless redesign, and each host adds rules of its own on top.
An MCP server is the program that lists a product's tools, such as search, read a record or create an order, so an AI agent can call them on a person's behalf.
A good first server has three to six well-described tools, reads before it writes, returns compact results, and uses OAuth 2.1 so the agent can never do more than the person it acts for.
An AI host is the application that runs MCP servers for a person, such as Claude, ChatGPT, VS Code, Cursor or Gemini CLI.
Each host adds its own rules to the MCP specification, on tool-name length, annotations, timeouts and result sizes, and the rules conflict. A server meant for several hosts is designed to the strictest of them.
Retrieval-augmented generation (RAG) answers questions with a language model by first retrieving the relevant pieces of your own material.
Most RAG systems split documents by token count and rely on vector similarity, which fails on structured material such as code, contracts and catalogues. Parsing the material into its real units and retrieving with more than one signal works better.
GraphRAG is retrieval-augmented generation that also follows the relationships between pieces of material, stored in a graph, rather than relying on text similarity alone.
In DevGuardian AI the codebase's dependencies are modelled in Neo4j, so an agent reviewing a change can ask what else breaks, not only what looks similar.
Server-side grounding is a pattern in which an AI agent returns only identifiers and the server builds everything the user sees from real tool results, so the agent cannot show an invented item.
We built it for MayAI, where a dish that is not on the menu or a price that is not real would be an order nobody can fulfil. It is a constraint rather than a prompt, so it holds when the model changes, and it can be tested.
Evals (evaluations) are repeatable tests of an AI system against a gold set of real questions or requests with known correct answers, scored before and after every change.
For an MCP server, an eval checks that an agent picks the right tool with the right arguments; for retrieval, that the right material comes back. Without one, nobody knows whether a change to a prompt, a model or a description helped.
Tell us what you are trying to do. We will come back with an honest take on what it would take.
Book an AI consultation