aka Frontier AI Agents

General Purpose, human driven

Overview

Agentic Coding Tools consist of the following 5 major areas:

  • AGENTS.md
  • Agent Skill
  • MCP
  • A2A
  • Subagent

Agentic Coding Tools diagram

AGENTS.md

AGENTS.md is what turns a LLM into an agent.

  • Can have multiple versions and take precedence closer to the files in the directory.
  • Has specific instruction on what to do when

Agent Skills

Agent Skills give the agent procedure - markdown instructions (plus optional scripts/resources) that get progressively loaded when relevant. A skill is where you encode how your org does the thing: ordering, guardrails, conventions, which tools to reach for and in what sequence.

  • The SKILL.MD file will have a description of when to invoke it when you want to do X
  • The composition pattern is: the skill references the MCP tools by name and orchestrates them.
  • Local filesystem ~/.agents/skills/

MCP

Model content Protocol (MCP) (Agent ↔︎ tool)

MCP is an open protocol for connecting AI application to tools, data sources, and workflows. MCP gives the agent Capabilities.

MCP Servers

The MCP server defines what the agent can do but says almost nothing about when and how to do it.

A2A

Agent 2 Agent (A2A) (Agent ↔︎ Agent)

A2A is an open standard for seamless communication and collaboration between AI agents.

Subagents

Create and use specialized AI subagents for task-specific workflows and improved context management.

  • A child agent that the main agent spawns to do a specific piece of work
  • Each sub agent runs in its own fresh context window.
  • Keeps the main agent context clean.
  • Common pattern but no formal standard documented behind them

Example of a subagent task:

“Go read 500 files and return to the main agent a summary of those files.”

Example subagents

  • Code reviewer
  • Debugger
  • Data Scientist
  • Business Analyst
  • Product Owner