Glossary

The AI-agent vocabulary is young and inconsistent. The same word means different things across vendors, and the most important distinctions -- plugin versus skill versus MCP, agent versus chatbot, vibe coding versus agentic engineering -- get blurred. This glossary is the canonical source for how Soleur uses each term. Every definition opens with a one-sentence quotable answer, expands in two to three sentences, and cites an external source.

Company-as-a-Service

Company-as-a-Service (CaaS) is a platform category where AI agents run every department of a business -- engineering, marketing, legal, finance, operations, product, sales, and support -- sharing one compounding knowledge base. Instead of hiring specialists or stitching together dozens of isolated SaaS tools, a founder installs one platform and gets a full AI organization. The structural difference from SaaS is cross-department memory: the marketing decisions inform the sales battlecards, and the legal audit references the approved privacy policy. Soleur coined and is built on this model -- see the Company-as-a-Service pillar. The broader shift it belongs to is documented in a16z's economic case for generative AI.

Agentic Engineering

Agentic engineering is a structured methodology where AI agents execute multi-step workflows under human oversight, compounding institutional knowledge across every session. It is the production-grade successor to vibe coding: it keeps the speed of AI generation but adds a plan before the code, tests and review after it, and a memory that carries the lesson forward. The academic framing of the split appears in arXiv 2505.19443. See the Agentic Engineering pillar.

AI Agent

An AI agent is a goal-directed software worker that takes an objective, plans the steps, uses tools, and returns a finished, inspectable result -- not a chatbot that merely answers questions. A useful agent has five properties: goal-oriented, tool-using, memory-carrying, verifiable, and correctable. LangChain frames it as a system that uses a language model to decide the control flow of an application. Read LangChain's definition of an agent and Anthropic's Building Effective Agents.

MCP (Model Context Protocol)

The Model Context Protocol (MCP) is an open standard that lets AI applications connect to external systems -- databases, APIs, file stores, SaaS products -- through a uniform interface. An MCP server is a separate process that exposes tools, resources, and prompts; it is how an agent reaches the world outside its own context. It answers the question "what systems can the model reach," as distinct from a skill's "what should the model know how to do." See the MCP announcement and modelcontextprotocol.io.

Claude Code Plugin

A Claude Code plugin is a distributable bundle that can contain commands, agents, skills, hooks, and MCP server configuration, installed as a single versioned unit and shared across a team via a marketplace. The plugin is the distribution layer; skills and MCP servers are the capability layers inside it. It answers "how do I ship a capability set to a team." Soleur itself is a plugin. See the Claude Code plugins documentation and the Claude Code plugins pillar.

Skill

A skill is a packaged unit of procedural knowledge -- a SKILL.md file plus optional scripts and references -- that teaches a model how to perform a task and loads on demand when relevant. Unlike an MCP server, a skill runs in-context with no separate process and no external connection; it encodes know-how, not reach. Skills can be bundled inside plugins and shared as libraries. See the Agent Skills documentation.

Knowledge Base (Compound Knowledge)

A knowledge base is the persistent, shared memory where an AI organization records decisions, patterns, and conventions so that every project starts ahead of the last one. Compounding knowledge is what separates an agentic organization from running many disconnected tools: a lesson learned once is written back and reused everywhere. Without it, every session re-learns the same lessons and the system gets slower with scale, not faster. See Knowledge Compounding in AI Development.

Human-in-the-Loop

Human-in-the-loop is a design pattern where a human reviews, approves, or corrects an AI system's actions at defined checkpoints, keeping the founder as the decision-maker and the agent as the executor. In agentic engineering it appears as the review checkpoint before merge and the correctability property of an agent. The discipline breaks when the human abandons that checkpoint. Anthropic's Building Effective Agents treats oversight as a first-class part of agent design.

Vibe Coding

Vibe coding is a conversational style of building software where you accept the code the model generates and run it to see if it works, prioritizing speed over structure. Andrej Karpathy coined the term in February 2025 to describe "fully giving in to the vibes." It excels at prototypes and weekend hacks and falls apart on the last 30% of production work -- integration, edge cases, maintainability. See Karpathy's original tweet and Addy Osmani's The 70% Problem.

Context Engineering

Context engineering is the practice of deciding what information enters a model's limited context window -- and what to leave out -- so the model has exactly what it needs to do the task well. Every loaded plugin, skill, and document costs context budget that competes with the founder's actual project. Good context engineering is why a focused knowledge base outperforms dumping everything into the prompt. See Anthropic's guidance in Building Effective Agents.

Where These Terms Come Together

These definitions are not academic. They describe one working system: a Company-as-a-Service platform, built with agentic engineering, shipped as a Claude Code plugin, made of agents and skills that share a knowledge base. Explore the 67 agents and 83 skills that put the vocabulary into practice.