Heidelberg AICurriculum

5 AI Agent Terms You Need to Know

IBM Technology ·2026-06-23 ·2 min read

Summary written by us from the video's transcript. The video, and everything in it, is IBM Technology's work.

Learn five core concepts—agents.md, agent skills, MCP, A2A, and sub‑agents—that enable modern AI agents to plan, code, and interact with tools and other agents.

Takeaways

  • Agents.md acts as a project‑specific README that directs an agent’s behavior and can be overridden by nested files.
  • Agent skills package task‑specific code and metadata in a folder, loading only when the request matches their description.
  • MCP standardizes how agents talk to external tools via an MCP server, eliminating custom connectors.
  • A2A enables seamless handoff between agents through shared agent cards, supporting coordinated workflows.
  • Sub‑agents let a parent agent offload large or parallelizable work into separate contexts, keeping the main context lightweight.

Agents.md: Project‑Specific Instructions

The **agents.md** file is a markdown document placed at the root of a project. When an agent starts work, it reads this file to learn project‑specific commands (e.g., test scripts), coding conventions, or PR title formats. Nested agents.md files can exist in sub‑directories; those closer to the working directory override higher‑level settings. OpenAI introduced the format and it is now maintained by the Agentic AI Foundation under the Linux foundation. Some platforms use a different name (e.g., Claude uses **Claude.md**) but the purpose is identical.

Agent Skills: On‑Demand Task Modules

An **agent skill** is a folder containing a **skill.md** file plus any scripts or resources needed for a specific task, such as building a PowerPoint deck. The markdown metadata describes when to invoke the skill (e.g., “invoke me when the user wants to create slides”). Skills are loaded only when relevant, keeping the agent’s context window lean. Like agents.md, skills follow an open standard supported by multiple platforms.

MCP – Model Context Protocol for Tool Access

**MCP** (Model Context Protocol) provides a uniform interface for agents to interact with external tools, databases, APIs, or SaaS services. An **MCP server** wraps the target resource and translates MCP calls into the native API (e.g., Notion or Stripe). This avoids building custom connectors for each integration. MCP originated at Anthropic and is now governed by the AAIF under the Linux foundation with broad industry backing.

A2A – Agent‑to‑Agent Communication

**A2A** (agent‑to‑agent) is an open protocol that lets agents coordinate without bespoke integrations. Each agent publishes an **agent card**, a description of its capabilities and communication method. Other agents can read the card and delegate work, such as a procurement agent handing off a contract to a finance approval agent. A2A was created by Google and is also hosted under the Linux foundation.

Sub‑Agents: Parallel and Scalable Workflows

When a single context window isn’t enough—e.g., reviewing thousands of code files—or when tasks can run in parallel, an agent can spawn **sub‑agents**. A sub‑agent is a child process with its own fresh context; it performs the assigned work (like reading 500 files) and returns the result to the parent. Multiple sub‑agents can operate concurrently, dramatically speeding up independent checks. Unlike the other terms, sub‑agents are a common pattern rather than a formal standard, but they appear consistently across modern agent systems.