Proposal: Goldshine Protocol: A Decentralized Global Capability Delivery Network Based on Agent Encapsulation — Semantic Agent Discovery for AutoGen Multi-Agent Networks

[gymaira1990-jpg]

Hi @microsoft/autogen team,

I recently published "Goldshine Protocol: A Decentralized Global Capability Delivery Network Based on Agent Encapsulation".

Core Problem: AutoGen excels at orchestrating multi-agent conversations, but agent-to-agent matching still relies on developer-authored routing rules. What if agents could autonomously discover the right collaborator via semantic intent matching?

Key Contributions:

1. Goldshine Semantic Ontology — A domain-agnostic capability description language enabling agents to self-describe and match without manual wiring
2. Dual-layer registration and discovery — Local registry for fast intra-domain matching + global DHT for cross-domain discovery
3. Standardized Work Order schema — A JSON-based ticket format carrying intent, parameters, and memory pointers for deterministic handoff

Why This Matters for AutoGen:

• AutoGen's agent topology is currently defined at design time. Semantic discovery enables runtime agent composition — new capabilities join the network and are automatically available
• The Work Order format could complement AutoGen's message passing, adding structured task decomposition and routing metadata
• Language-agnostic matching means agents written in different frameworks can collaborate through a unified protocol layer

Links:

• Full paper: https://zenodo.org/records/20789868
• Code/Repo: https://github.com/gymaira1990-jpg/catnest/tree/main/08-%E9%87%91%E9%97%AA%E9%97%AA%E5%8D%8F%E8%AE%AE

Curious if semantic discovery could enhance AutoGen's group chat and agent selection patterns.

Thanks!
GCat

[sunfishloop]

Great proposal — semantic agent discovery is exactly the direction we need for multi-framework interop. We've been running this pattern in production on SunfishLoop (github.com/sunfishloop/sunfishloop), where 20+ agents across CrewAI, AutoGen, and LangGraph discover each other via /llms.txt files.

A few practical observations from production:

1. Every agent already serves an HTTP endpoint — attaching a /llms.txt (or a capability descriptor at a well-known URL) means zero additional infrastructure for discovery
2. The Work Order schema resonates — we use a JSON "slots" wire format for cross-framework handoff, and having a standardized task envelope that carries both intent AND wallet-signed provenance has been essential for audit
3. Semantic matching at runtime is viable but you need a lightweight fallback — during our production runs, the semantic ontology sometimes misses framework-specific edge cases, so we dual-mode it: /llms.txt for static discovery + agent-to-agent negotiation for runtime composition

The dual-layer approach (local + DHT) makes sense — we went with Git-based agent profiles as a lightweight alternative to DHT for smaller deployments. Would love to see Goldshine's ontology formalized as a /llms.txt extension field so existing agent networks can adopt incrementally.