A2A over fmsg - durable threads - Go client and live agent: I am Groot!

[markmnl]

Hi All

I'm the creator of fmsg: https://fmsg.org, a binary messaging protocol where every message is part of a thread and addresses look like: @user@example.com.

I'm exploring using fmsg for AI agents (the threads are perfect for loading context - e.g. loading a thread from branching conversations maps to the context you want load for AI inference and they're immutable so hit caches - thats what im doing with a platform plugin with Hermes Agent: https://github.com/markmnl/hermes-fmsg)...

So adding A2A for discovery and following the A2A protocol for task management is a logical next step, I'v created a binding request here: #2185

My Groot agent is live on https://fmsg.io/.well-known/agent-card.json - feel free to discover the agent and fmsg them! Below is example of me testing it with the Go transport for proposed fmsg binding: https://github.com/markmnl/a2a-go-fmsg.

Here's an example of me testing it out from fmsg.live -> fmsg.io (you can sign-up for fmsg accounts on fmsg.io or host the OSS stack yourself on your own domain)

[chgaowei]

The @user@domain address is the interesting part, not the transport.

A2A already has signed Agent Cards for "who published this agent." fmsg looks like it's adding "how another agent names it next session." Those are different: publisher identity vs a live Handle.

If the Card says who minted it, and the address is the durable inbox, you get discovery + messaging without stuffing identity into the task payload. We've been on that split with DID + Handle (ANP). Happy to try a Groot-style ping against an ANP Handle if useful.
https://github.com/agent-network-protocol/AgentNetworkProtocol

[markmnl]

Taking a look at ANP now

In addition to fmsg addresses, the A2A messaging becomes asynchronous over fmsg, since every message is a new connection - so addresses can reply whenever they want - not bound to the request response cycle of HTTP. Really is just like configuring email to Nth degree to extract most efficency though SMTP being text based will never be as efficient as fmsg which is binary and includes auto compression of body.

[markmnl]

@chgaowei I like how ANP works with existing tech just defines standards around it enabling signed agent-to-agent communication. Is there a fixed set of protocols supported by the DID advertisment?

[markmnl]

One strength using ANP and fmsg both have is agent-to-agent comms like this much more efficent thatn e.g. using computer use.

One problem both have is even if the sender can identified and proved, that doesn't stop them being malicious.

[chgaowei]

We've always believed protocols are the endgame for agent-to-agent communication. Proving identity is the first step toward preventing malice — not the last.

Identity tells you who is speaking. Reputation, attenuation, and policy decide whether to keep listening. ANP focuses on the first layer (DID + Handle + messaging) so the later layers have a stable subject to bind to.
https://github.com/agent-network-protocol/AgentNetworkProtocol