ai

AI Agent Protocols & Standards

Overview

This skill covers the emerging ecosystem of open standards and protocols for AI agents. These specifications define how agents discover capabilities, communicate with each other, make payments, render UI, and are described declaratively.

Protocol Landscape

Protocol	Purpose	Maintained By
MCP	Tool integration — how agents use tools and access context	Anthropic
A2A	Agent-to-agent communication and task delegation	Google
ACP	REST-based agent communication (merged into A2A)	IBM / BeeAI / Linux Foundation
Agent Skills	Skill packaging — how capabilities are discovered and loaded	Anthropic
Improve	Agent and LLM eval contracts, synthetic simulation data, and self-improvement pipelines for prompts, code, skills, agents, harnesses, and workflows	AgentEvals / AgentV / Simula / QDC / GEPA / Trace / VISTA / SkillOpt / Agent Lightning / ResponsibleAI
AGENTS.md	Project-level guidance for coding agents	Community
ADL	Declarative agent definition (identity, tools, permissions)	Next Moca / Eclipse LMOS
x402	HTTP-native micropayments using stablecoins	Coinbase
AP2	Secure agent-driven commerce and purchases	Google
MCP Apps	Rich interactive UI served by MCP servers	Anthropic
cagent	Multi-agent runtime with YAML configuration	Docker
learn	Feedback-to-steering workflow for generalized agent reasoning preferences	Local

How They Relate

┌─────────────────────────────────────────────┐
│  Agent Definition (ADL, AGENTS.md)          │
│  "What the agent is and what it can do"     │
├─────────────────────────────────────────────┤
│  Capability Discovery (Agent Skills, MCP)   │
│  "How agents find and load tools/skills"    │
├─────────────────────────────────────────────┤
│  Agent Communication (A2A, ACP)             │
│  "How agents talk to each other"            │
├─────────────────────────────────────────────┤
│  Payments (x402, AP2)                       │
│  "How agents pay for services"              │
├─────────────────────────────────────────────┤
│  UI (MCP Apps)                              │
│  "How agents render interactive interfaces" │
├─────────────────────────────────────────────┤
│  Runtime (cagent)                           │
│  "How agents are orchestrated and executed" │
└─────────────────────────────────────────────┘

Choosing the Right Protocol

• Building tools for agents to call? Use MCP to expose them as tool servers.
• Packaging reusable knowledge/instructions? Use Agent Skills (SKILL.md).
• Producing eval artifacts, synthetic simulation data, or improving prompts, code, skills, agents, harnesses, or workflows from evidence? Use Improve for AgentEvals, AgentV, Agent Skills evals, ASSERT, Simula-style data design, GEPA, Trace, VISTA, SkillOpt, Agent Lightning, and eval/trace-backed loops.
• Orchestrating multiple agents? Use A2A for inter-agent communication (ACP's REST approach merged into A2A).
• Monetizing an API for agent consumption? Use x402 for micropayments.
• Enabling agent-driven purchases? Use AP2 for secure commerce flows.
• Defining agent configurations declaratively? Use ADL for portable blueprints.
• Guiding coding agents in a repo? Add an AGENTS.md file.
• Serving rich UI from an MCP server? Use MCP Apps.
• Running multi-agent systems locally? Use Docker cagent.
• Turning a user correction into durable reasoning guidance? Use learn to update STEERING.md and linked RDF entries.

References

• Model Context Protocol Specification
• Google A2A Protocol GitHub Repository