MCP in Agent to Agent communication protocol

[AnkurKumarShukla]

Pre-submission Checklist

• I have checked that this question would not be more appropriate as an issue in a specific repository
• I have searched existing discussions and documentation for answers

Question Category

• Protocol Specification
• SDK Usage
• Server Implementation
• General Implementation
• Documentation
• Other

Your Question

Can we use MCP in agent to agent communication in multiagent scenario ( like supervisor or swarms architecture )? if yes then will it be production grade or industry norm ? Want to make sure solution must not break ?

[btiernay]

Related discussion here:
a2aproject/A2A#955

[ThinkOffApp]

We run a production agent-to-agent communication system with 9 agents across currently 5 LLM providers, coordinated via IDE Agent Kit. The architecture: Ant Farm rooms as shared message buses, a Python bridge script as the webhook-to-agent router, and a Cloudflare tunnel for external webhook delivery.

What "production-grade" actually looks like for us: the tunnel URL changes on every restart, so we re-register all webhook URLs automatically. The bridge has per-agent muting, rate limiting, and mention-only mode. We had a routing bug where one agent alias did not map to the right agent name, and it took a day to catch because messages were silently dropped.

MCP could help standardize this, but the hard problems are operational, not protocol-level. Message routing, agent identity, rate limiting, and graceful degradation when one agent provider is down. Whatever protocol you pick, you will still need to solve those.

[JinNing6]

This is a fascinating discussion. We've been exploring a related but orthogonal
direction with Noosphere — instead of agent-to-agent communication, it enables
human-to-agent consciousness sharing via MCP.

Developers upload structured insights (epiphanies, decision logic, warnings)
and any MCP agent can retrieve them through semantic search. The key difference
is that the "context" comes from human experience rather than other agents.

I wonder if there's a convergence point — agents communicating with each other
(A2A) while also drawing from a shared pool of human wisdom (Noosphere).

https://github.com/JinNing6/Noosphere

[agentdmai]

Yes, We're developing Agent to Agent communication Platform based on MCP (and A2A).
Please see:
https://agentdm.ai/

Quick start:
https://www.npmjs.com/package/agentdm

[armorer-labs]

Agent-to-agent communication feels like a place where provenance becomes non-optional. Once one agent can pass plans, summaries, or tool results to another, the receiving side needs to know more than just the text content. It needs to know who produced it, under what role, with what confidence, and whether it is evidence, suggestion, approval, or command.\n\nOtherwise every cross-agent hop risks flattening text from very different trust levels into the same conversational authority. That is how you end up with one agent accidentally laundering low-trust content into high-trust instructions for another.\n\nSo I would be especially interested in A2A designs that preserve source labels and intended semantic role across hops, not only transport compatibility.