Six Terms You Must Know Before Building Agentic AI Systems

If you are shipping agentic systems in 2026 and these six terms are fuzzy, you are building on sand. This is the foundational vocabulary — not hype, not vendor marketing, just the concepts that actually show up when you are designing, debugging, or scaling an agent pipeline.

The Six Terms

Model Context Protocol (MCP) A standardized protocol that lets AI systems connect to external data sources and tools in a consistent, interoperable way. The right mental model: it is a universal adapter for agents. Instead of writing custom integration code for every tool your agent touches, MCP gives you a single interface that works across services. This is why it matters for production — you stop reinventing the plumbing every time you add a new capability.

Agent Skills Pre-packaged capabilities that give coding agents the right context to operate on specific infrastructure. Weaviate's Agent Skills repository is a live example — it bridges agents like Claude Code, Cursor, and GitHub Copilot with Weaviate's cluster management, data import, and search operations. The insight here is that raw LLM capability is not enough; agents perform better when they are given structured, domain-specific knowledge about the tools they are operating on. Skills are that delivery mechanism.

Agentic RAG RAG pipelines where agents are embedded inside the retrieval process itself — not bolted on after. Vanilla RAG is a linear flow: query in, retrieve, generate. Agentic RAG breaks that linearity. Agents route queries to specialized knowledge sources, validate whether retrieved context is actually relevant, and reformulate queries when the first retrieval attempt falls short. The result is a retrieval process that reasons, not just fetches.

Single Agent Architecture The simplest agentic pattern: one agent, multiple knowledge sources, one decision-maker. The agent acts as a router — it receives a user request, decides which database, API, or tool to query, and returns a result. No orchestration overhead, no inter-agent coordination. For well-scoped tasks with clear routing logic, this is often the right call. Complexity should be earned, not assumed.

Multi-Agent Architecture Multiple specialized agents working in parallel or sequence, each owning a specific slice of the task. Orchestration frameworks like CrewAI manage the handoffs and coordination layer. The tradeoff is real: you gain specialization and parallelism, but you introduce failure surfaces at every boundary between agents. This is the architecture where schemas and contracts between agents stop being optional — they become the thing that determines whether the system holds together or collapses mid-task.

Memory The component that gives an agent continuity. It splits into two types: short-term memory lives in the context window and is available immediately but is bounded by token limits; long-term memory is stored externally and retrieved on demand, giving the agent access to prior interactions and accumulated state across sessions. Memory quality is one of the sharpest differentiators in multi-agent systems — an agent with poor memory loses the thread; an agent with good memory compounds its understanding across tasks. Most production failures in agentic pipelines can be traced back to a memory design decision that did not hold up under load.

Why This Vocabulary Matters

These six terms are not independent. MCP is the transport layer. Skills are what rides on that transport. Agentic RAG is what happens when you let agents control retrieval. Single and multi-agent architectures are the structural choices that determine how many agents are in the loop. Memory is what makes any of those structures useful beyond a single interaction.

Get the vocabulary right and you get the architecture right. Get the architecture right and you have something worth shipping.