Google cloud launches Agent2Agent protocol
Google has launched a new protocol spec for agents to talk to each other. Link to the spec. Homepage here.
From the blog:
“Today, we’re launching a new, open protocol called Agent2Agent (A2A), with support and contributions from more than 50 technology partners like 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 will allow AI agents to communicate with each other, securely exchange information, and coordinate actions on top of various enterprise platforms or applications. We believe the A2A framework will add significant value for customers, whose AI agents will now be able to work across their entire enterprise application estates.”
There is a ton of consulting firms as design partners here. It suggests to me that this entire thing was thought out from a POV of automating knowledge work which is a good sign.
“A2A is an open protocol enabling seamless agent collaboration across different frameworks and vendors. It adheres to five principles: embracing natural agent interactions without forcing agents into mere “tool” roles; building on established standards like HTTP and JSON-RPC for easy integration; prioritizing enterprise-grade security; supporting tasks ranging from quick actions to extended research with real-time feedback; and accommodating multiple modalities including text, audio, and video streaming.”
Even though Google claims MCP is designed for agent-to-tools interaction and that A2A is better for agent-to-agent, there’s nothing preventing MCP from being used for agent-to-agent too, as the distinction between agents and tools is rather arbitrary.
In terms of comparison to MCP:
- MCP - exposes prompts, resources and tools to a host, who can do whatever they like
- A2A - exposes capability discovery, tasks, collaboration/chat, user experience discussions (can we embed an image or or a website?).
High-level it makes sense to agree on these concepts. I just wonder if we really need a fully specified protocol? I personally think that this could have been an addon for MCP instead of an entire new spec itself. After reviewing the A2A spec and sample implementations, here’s what I think:
- A2A appears to supersede MCP with more comprehensive features - implementing A2A likely eliminates the need for MCP.
- A2A addresses MCP pain points with practical solutions:
- Structured data transmission protocols
- Token-based security for function calls
- DNS-based discovery and security
The UI demos require more sophisticated clients with iframe and reconnection capabilities. Text-only clients exist but UI negotiation and non-HTML client support remain unclear.
On the other hand, they also published something called the “Agentcard” which should make discovery easier for these agents and we probably won’t see one more explosion MCP server directory type thing happening again.
Parallely, they also launched a full stack agent dev kit. Built in primitives for doing agent orchestration, connections to pre-built tools, eval harness and ofcourse easy deployment to GCP. Its a direction competition to Langchain, CrewAI and similar platforms. Interestingly Langchain is a tech partner for developing A2A.