Can AI agents talk to other AI agents?

What would it mean for agents to talk to each other?

Picture a customer asking their own AI assistant to find a plumber who can come on Thursday. For that assistant to ask a plumbing company's agent directly, several things have to work: the two agents need to find each other, learn what each can do, prove who they are, agree on what each is allowed to do, and use a shared message format.

Protocols handle the shared format and some of the discovery and security plumbing. They don't decide whether a business's agent is accurate or trustworthy, and they don't make any particular assistant use them.

What is the Model Context Protocol (MCP)?

The Model Context Protocol is an open-source standard for connecting AI applications to external systems, such as data sources, tools and workflows. Its documentation compares MCP to a USB-C port for AI applications. ^[1]

The MCP documentation lists AI assistants such as Claude and ChatGPT, and development tools such as Visual Studio Code and Cursor, among the products that support it. ^[1] In MCP terms, a host is the AI application, a client is its connector, and a server provides context and capabilities such as resources, prompts and tools. ^[2]

MCP is about an agent using tools, not about two agents conversing. The A2A project describes the split this way: MCP defines how an AI agent uses individual tools and resources, such as a database or an API. ^[7]

What is the Agent2Agent (A2A) protocol?

Google announced the Agent2Agent protocol on April 9, 2025, with support from more than 50 technology partners, describing it as complementary to Anthropic's Model Context Protocol and built on existing standards including HTTP, SSE and JSON-RPC. ^[4]

A2A was originally developed by Google and donated to the Linux Foundation. Its technical steering committee includes AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow. ^[5]

A2A is designed so agents can collaborate without exposing their internal logic, memory or proprietary tools, and it supports long-running tasks. ^[6] The specification site lists 1.0.0 as its latest released version. ^[8]

The repair-shop example comes from the A2A documentation, which describes A2A as handling higher-level, conversational, task-oriented interactions between agents, while MCP lets the shop's own agent use its tools. ^[7]

How do agents find each other and learn what the other can do?

In A2A, an agent publishes an Agent Card: a JSON metadata document describing its identity, capabilities, skills, service endpoint and authentication requirements. The specification gives a standard address for it under the agent's domain, at /.well-known/agent-card.json. ^[8]

A client agent fetches the other agent's Agent Card to learn its capabilities and endpoint before sending it work. ^[6] An Agent Card is a self-description, so it tells you what an agent claims to do, not how well it does it.

How do identity and permissions work between agents?

A2A follows standard web practices for authentication, authorization, security, privacy, tracing and monitoring. ^[6] The specification leaves authorization logic to each implementation, allows Agent Cards to be digitally signed, and says clients should verify a signature before trusting a card. ^[8]

For MCP, authorization is optional. When it is used over HTTP, it is based on OAuth 2.1 and related standards. ^[3] The MCP specification says users must explicitly consent to data access and tool use, and states that MCP itself cannot enforce these principles at the protocol level. ^[2]

The practical point for a business: a protocol makes identity and permissions possible to check, but the people who build and operate each agent decide how carefully they are checked. Proving that an agent controls a domain is not the same as proving the business behind it is reputable.

What is demonstrated today, and what isn't?

It helps to separate three layers.

Demonstrated today

MCP is supported by widely used AI assistants and developer tools. ^[1] A2A has a published, versioned specification ^[8] and is governed under the Linux Foundation with a multi-company steering committee. ^[5] Developers can build agents today that publish Agent Cards and exchange A2A messages.

Still emerging

Shared directories for finding agents, ways to judge an unfamiliar agent's reliability, and consistent permission models across companies are still forming. Whether a particular consumer assistant will contact an outside business's agent depends on that product's own documentation, and you shouldn't assume it does.

Future possibilities

A customer's assistant checking a local business's availability, requesting a quote, or arranging a visit through the business's own agent is technically plausible with these protocols. Nobody can promise when, or whether, that becomes common for small businesses.

Where does Bryno fit?

Bryno does not run an agent network. Its AI agent lives on a person's or business's digital card and answers the people who visit that card.

Agent-to-agent messages

Bryno's AI agent answers people who visit a card; it does not accept messages from other AI agents. For what a business can usefully do now, see what an AI business agent is and how AI assistants find local businesses.