Checking my survey's description of Agent2Agent 1.0

[jwade83]

Hello,

I'm an independent researcher writing a survey of agent-native systems: the protocols, message buses, memory layers and operating-system support that multi-agent setups now rely on. I'm using it to design a small multi-agent lab at home, and before it shapes that design I'm checking each section with the people who build or wrote the thing it describes.

Here is how the survey currently describes Agent2Agent:

• Version 1.0 (12 March 2026) separated the protocol from its transport bindings (JSON-RPC, gRPC, REST), added task listing with filters, and replaced the implicit and password OAuth flows with device-code and PKCE.
• Version 1.0.1 followed on 26 May 2026 as a fixes release.
• Signed agent cards have been available since version 0.3 (July 2025).
• Agent2Agent joined the Agentic AI Foundation in August 2026, alongside Model Context Protocol.
• The survey recommends it for agent-to-agent traffic between machines on a home network, with agent cards doubling as local discovery.

My questions:

1. Is Agent2Agent a reasonable fit for a small private deployment (a handful of agents across two or three machines on one network), or is it designed mainly for traffic between organisations, with something lighter expected inside one network?
2. Is publishing and finding agent cards on a local network (for example alongside multicast DNS) a supported or anticipated pattern?
3. Is anything on the roadmap after 1.0.1 likely to change the points above, such as requiring signed cards?

A one-line "that's right" helps; corrections or caveats help more. If it's alright, I'll link this thread from the survey's sources so the correction travels with the claim.

I drafted this post with help from an AI assistant; the questions and the lab are my own.

Thanks,
John Wade

[capitaharlock]

Not a maintainer, but one small thing on the wording. As far as I know the spec doesn't define discovery on a local network, a client fetches the card from a well-known URL or gets it configured. So "agent cards doubling as local discovery" reads like A2A does that itself.

Phrasing the mDNS part as your own setup on top of A2A would keep it accurate, and the maintainers can say whether it's a pattern they expect.