AI agents from multiple sources need more than the bare-bones Model Context Protocol (MCP) to reliably work together and two new standards are developing to ensure they do: A2A and AAIF.
The Agent-to-agent (A2A) and Agentic AI Foundation (AAIF) both reflect a recognition by players in the AI agent ecosystem that they need multi-level standards defining how agents from different vendors interoperate. As virtually all storage software suppliers are developing agents or ways of working with agents, these initiatives are worth watching and tracking.
We’d better agree on what constitutes an AI agent. IBM says: “An artificial intelligence (AI) agent is a system that autonomously performs tasks by designing workflows with available tools. AI agents can encompass a wide range of functions beyond natural language processing including decision-making, problem-solving, interacting with external environments, and performing actions.
“AI agents solve complex tasks across enterprise applications, including software design, IT automation, code generation and conversational assistance. They use the advanced natural language processing techniques of large language models (LLMs) to comprehend and respond to user inputs step-by-step and determine when to call on external tools.”
Google announced its A2A agent interoperability protocol in April. It wanted to avoid the need for multiple specific and customized point-to-point agent-to-agent communications methods. A2A is designed so that agents “interoperate with each other, even if they were built by different vendors or in a different framework.” Google was supported by more than 50 partners including “Atlassian, Box, Cohere, Intuit, Langchain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, UKG and Workday, and leading service providers including Accenture, BCG, Capgemini, Cognizant, Deloitte, HCLTech, Infosys, KPMG, McKinsey, PwC, TCS, and Wipro.”
The A2A protocol starts with client and remote agents. The client agent sets up a task with a defined output, known as an “artifact,” and requests the remote agent to carry it out. A2A assumes the agents may not share memory, tools, or context. Tasks can be short – seconds or minutes – or long-lived – hours or days – with status messages relaying their state. A client agent won’t necessarily know what remote agents are available for a task so remote agents can publicize their capabilities using a JSON format Agent Card. This enables the discovery of remote agent functionality.
The communications or messages between the agents are based on existing protocols, such as HTTP, SSE (Server-Sent Events), and JSON-RPC. A message will have formal “parts” with a defined content type.
Naturally, there is a security aspect to this with enterprise-level authentication, authorization, privacy, tracking, and monitoring.
Google donated the A2A protocol to the Linux Foundation and there is an A2A Protocol website. This contains an example of a foreign travel AI assistant carrying out a chain of tasks when a user prompts it to arrange a trip:
The AI assistant acts as the client or orchestrator, and requests the four remote agents – flight, hotel, currency, and tours – to carry out their specific tasks. It then returns a completed travel plan to the user. We’re told agents can reason, plan, and delegate tasks to other agents. You can study an A2A request lifecycle sequence on the site to understand API operational flows in an A2A request.
A four-level agent stack is defined: A2A protocol, Vertex AI Engine, MCP, and ADK.
Agents can use MCP to communicate with non-agent entities, such as MinIO’s AIStor object storage software and Snowflake’s data warehouse. A2A refers to these entities as tools.
Google has developed the open-source and modular Agent Development Kit (ADK) framework.
Google’s A2A specification draft can be read here. A production-ready version of the protocol was hoped to be available by the end of 2025 but that looks unlikely now.
AAIF
The Agentic AI Foundation (AAIF) was co-founded by OpenAI, Anthropic, and Block under the Linux Foundation with support from Google, Microsoft, AWS, Bloomberg, and Cloudflare. It aims to provide “neutral stewardship for open, interoperable infrastructure as agentic AI systems move from experimentation into real-world production.”
OpenAI says the AAIF is intended to be “a neutral home where agent interoperability standards can be developed, governed, and extended collaboratively.”
OpenAI contributed its AGENTS.md format to the AAIF, providing agents with project-specific instructions and context, to ensure, it said, long-term support and adoption of this open format across the community.
OpenAI says: “Since its release in August 2025, AGENTS.md has been adopted by more than 60,000 open source projects and agent frameworks including Amp, Codex, Cursor, Devin, Factory, Gemini CLI, Github Copilot, Jules and VS Code among others – reflecting growing alignment around shared, vendor-neutral conventions as agents enter production.”
Anthropic has contributed MCP to the AAIF, and Block has donated its Goose open-source AI agent framework for connecting large language models (LLMs) to activities and tools.
On joining the AAIF, Snowflake said it believes “that standardization of agentic AI tooling, driven by neutral collaboration in open source communities, is the best way to ensure interoperability and prevent fragmentation in this rapidly evolving space.”
There are four levels of membership: Platinum, Gold, Silver, and Associate. Platinum-level membership costs $350,000 and includes Anthropic, AWS, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI. The Platinum list is closed and the eight contributed $2.6 million in total.
Gold membership costs $200,000 and the list includes Cisco, Datadog, IBM, Okta, Oracle, Salesforce, Snowflake, and others – 18 in total and contributing $3.6 million. Silver membership costs $10,000, but rises with employee count to $95,000 for 5,000-plus staff. Its list features Elasticsearch, Hugging Face, Mirantis, SUSE, Uber, and many more – 23 in total. Say the average is $35,000 per company, that gives us $805,000. That’s the supplier entry level. Associate Membership, which is free, is restricted to approved non-profits, academics, and government bodies. The Linux Foundation has likely taken in $7 million overall as a result of the AAIF being set up.
Will the Linux Foundation combine or link A2A and the AAIF? It seems an obvious step. We shall see.
