I and two other colleagues have been building Summoner, a decorator-first Python SDK with a Rust relay for live, duplex agent-to-agent networking across machines. Basically, think MMO but for AI agents.

It's currently in early beta (v1.0) (expect a few sharp edges), but we have already built 25 template agents from which you can build your projects (hello world, question/answer, protocol handshakes, negotiation, etc.) ;

### What it is

- A tiny runtime where you write @receive / @send handlers and let the relay move messages between agents in real time. - Routes are simple strings that behave like labeled edges; the runtime compiles them into a small automaton so agents can coordinate without hand-building graphs.

### Why it’s different

1) Not a model-to-tool connector like MCP: Summoner focuses on agent-to-agent sessions and orchestration, not model-to-tool calls.

2) Not server-anchored agent executors like A2A: Summoner agents are mobile and duplex by default, so they can initiate/serve and move between relays without being wrapped as server executors.

3) Compared to LangChain/LangGraph, Summoner aims for less ceremony: you don't draw the graph, you just register handlers (MCP-like) and let routes drive the flow.

### What I'm looking for

- Critical feedback on the SDK (see https://github.com/Summoner-Network/summoner-docs)

- Where would you want MCP tools or A2A discovery layered in?

- Reliability/perf corners you would test first

- Security review of the state-machine approach

### Links

- Examples (start here): https://github.com/Summoner-Network/summoner-agents

- Docs / design notes: https://github.com/Summoner-Network/summoner-docs

- Core runtime: https://github.com/Summoner-Network/summoner-core

- Project page: https://summoner.org