A short, neutral glossary of the terms that come up when companies start working with AI agents — plain English, no hype, with a link to the service that solves the underlying problem.
An AI agent is a program that takes a goal, plans steps, and uses tools (web, APIs, files) to reach the goal without needing a human to approve every micro-step.
Matters when you want AI to do a whole job end-to-end, not just answer a single question.
MCP is an open standard for how AI agents connect to tools and data sources. Think USB-C for LLMs: a server exposes tools, a client (the agent) calls them.
Matters when you connect AI to internal systems and want to avoid lock-in to a single model vendor’s API.
A sandbox is an isolated environment where an AI agent can run tools and code without being able to touch production or sensitive data.
Matters the moment an agent will write, execute, or do anything irreversible — sandbox first, production later.
Provenance is being able to trace where an AI-generated answer came from: which source, which model, which prompt, which version.
Matters when someone needs to trust, audit, or revise an AI output — especially in regulated industries.
A connector is the glue between an AI agent and an external system (CRM, calendar, database). It translates the agent’s tool calls into the system’s API.
Matters when AI needs to act inside an existing system, not just talk about it.
The workflow owner is the person who owns a concrete workflow in the business and can describe what actually happens step by step — not the manager, not IT.
Matters in every AI project: without a workflow owner you build solutions for a process nobody actually runs.
Tool Forge is when you build small, specialized AI tools on top of a general model — shaped to fit your specific way of working.
Matters when general-purpose chatbots aren’t enough and you need something shaped to your process — without building a whole product.
An evaluation loop is a recurring routine where you measure whether the AI solution still gives the right answer — and adjust prompt, data, or model when it drifts.
Matters once AI is in production: without an evaluation loop nobody notices quality drops until customers do.