Navigating Agent AI Standards: Foundations, Protocols, Competition, and Path to Convergence

As an AI and Agentic AI practitioner and enthusiast, I have watched the agentic AI space undergo a massive structural shift over the past year. We have moved rapidly past single LLM prompt wrappers into the era of autonomous, multi-agent systems. However, with this explosive growth has come an inevitable wave of standardizations, acronyms, and governing foundations. If you look at the landscape today, you are confronted with a plethora of open-source initiatives and projects: Agentic AI Foundation, AGNTCY, Linux Foundation-hosted projects such as A2A, BeeAI, Agent Skills by Anthropic, and new entrants like Agent Plugins by Google.

This, naturally, brings up a few questions:

• Are these standards competing with each other?
• Which ones should my team adopt today?

In this post, I want to step back and provide a clear, practitioner-focused guide to the entire Agentic AI standards ecosystem. We will explore the governing foundations, map out each protocol layer, analyze where they compete versus complement each other, and outline recommendations for how these open initiatives can converge into a unified “Internet of Agents.”

Standards

To make sense of the ecosystem, we must first look at who governs what. Standards in the AI agent world are driven by neutral foundations (primarily the Linux Foundation) alongside cross-industry technical steering committees.

Key Foundations & Alliances

Let us map out the open standards, specifications, and projects hosted under each foundation or driven by key industry vendors.

Agentic AI Foundation (AAIF)

The AAIF was established under the Linux Foundation to host foundational projects for agent tool interaction, developer context, and data plane proxying:

• Model Context Protocol (MCP): Donated by Anthropic. A universal standard for connecting AI models to data sources, local files, and external tools (mcp.json).
• Goose: Donated by Block. An open-source, local-first AI agent execution framework.
• AGENTS.md: Donated by OpenAI. A standardized file format providing repository-level instructions to coding agents.
• agentgateway: Donated by Solo.io and community maintainers. An open-source, high-performance Rust-based AI-native proxy/gateway built specifically to route MCP, A2A, and LLM inference traffic at the network edge.

Linux Foundation

• AGNTCY: I have written about this earlier as a series of articles. Donated by Cisco with founding ecosystem partners including Dell Technologies, Google Cloud, Oracle, Red Hat, LangChain, and LlamaIndex, AGNTCY provides a complete horizontal infrastructure stack for multi-agent discovery, encrypted transport, identity, and invocation:

  • OASF (Open Agentic Schema Framework): A standardized, attribute-based taxonomy separating root actions (skills) from context (domains) with numeric IDs.

  • Agent Directory Service (ADS): A federated, peer-to-peer Kademlia Distributed Hash Table (DHT) for capability-based agent discovery.

  • SLIM (Secure Low-Latency Interactive Messaging): High-throughput gRPC transport with RFC 9420 MLS end-to-end encryption and relay gateways.

  • Identity & Trust: W3C Decentralized Identifiers (DIDs), Verifiable Credentials, and Tool-Based Access Control (TBAC).

  • ACP (Agent Connect Protocol): An OpenAPI 3.1.1 compliant REST API specification for stateful Thread management and Run execution.

• A2A Protocol: Originally introduced by Google and ecosystem partners, A2A is an open specification for peer-to-peer agent messaging:

  • A2A Wire Protocol: Defines task lifecycle message envelopes (AgentRequest, AgentResponse).

  • AgentCard Spec: A JSON metadata descriptor specifying HTTP web service endpoints, supported input/output media modes, and OAuth2 security schemes.

Cross-Industry Initiatives

Created by maintainers from leading AI labs and innovators:

• Agent Resource Discovery (ARD / ai-catalog.json): Backed by an industry coalition (Google, Microsoft, Cisco, NVIDIA, Hugging Face, Snowflake). Defines a domain-anchored manifest format (/.well-known/ai-catalog.json) for publishing local domain agent and tool catalogs.

• Agent Plugins Specification: A portable package format (plugin.json) that bundles Agent Skills (skills/SKILL.md) and MCP servers (mcp.json) into a single, vendor-neutral directory structure compatible across different IDEs and agent clients.

• Agent Skills Specification: A directory format featuring a mandatory SKILL.mdfile (YAML frontmatter + Markdown steps). Uses progressive disclosure so agents scan lightweight metadata at boot and load full instructions/scripts into the context window only when a task matches.

Open Knowledge Format (OKF)

Created by Google and available as an open specification, OKF provides an open, human, and agent-friendly format for representing knowledge provided to the agent. It is designed to be authored by people, generated by agents, exchanged across organizations, and consumed by both.

Competing vs. Complementing Dynamics

Are these projects competing, or do they fit together? The answer depends on which part of the stack you examine. When viewed holistically, these standards form a remarkably cohesive operational flow:

1. Discovery (ARD + ADS): Domain hosts publish ai-catalog.json (ARD), while federated networks publish OASF records into the P2P Directory (ADS).
2. Packaging: Agent Plugins bundles MCP servers (mcp.json) and Agent Skills (SKILL.md).
3. Ingress & Transport: agentgatewayhandles edge rate-limiting, proxying traffic into AGNTCY’s ACP Invocation Gateway over SLIM RFC 9420 MLS encrypted channels.

Despite this synergy, there are genuine areas of overlap that require coordination:

Recommendations for Convergence

To ensure the multi-agent ecosystem moves toward seamless interoperability rather than fragmentation, I recommend 4 concrete steps for convergence:

1. Standardize “ARD-over-ADS” so domain ai-catalog.json files are automatically indexed into the peer-to-peer Kademlia DHT.
2. Allow A2A AgentCard URLs and agent-plugins.org CIDs to serve as canonical artifact locators inside OASF records.
3. Configure agentgateway to natively proxy ACP routes (/threads) and carry A2A message envelopes over SLIM encrypted channels.
4. Align AAIF, AGNTCY, agentgateway, and Agent Plugins steering committees around shared taxonomy definitions and validation APIs.

Developers and enterprises implementing autonomous multi-agent systems must understand and carefully evaluate the competing and complementing dynamics of different open standards. How is your team structuring your AI agent stack today? Are you adopting MCP, Agent Plugins, or building on AGNTCY?