A2A Agent Registry and Discovery
1. Registry and Discovery over A2A
Should we build an Agent Registry and Discovery Agent that allows agents to register and discover other agents over the A2A protocol, much like DNS?
For this purpose, we should use A2A rather than MCP, because agents should register and discover one another symmetrically as agents, not through an asymmetric client-server tool model.
2. Semantic Discovery with RAG
When an agent registers itself, the registry should create a RAG index from its JSON metadata, including descriptions of its tools and capabilities, using a system such as Azure AI Search.
This would allow an agent to find other agents by describing the required capability in natural language, rather than relying on exact names or predefined keywords.
3. Trust and Authorization
When agents register their capabilities, the registry should also maintain a trust parameter. Otherwise, agents may use capabilities provided by untrusted agents.
The registry should also issue or broker tokens to requesting agents. Otherwise, agents could connect without sufficient control to agents that access sensitive or important data.
4. A2A-Native Cloud Microservices
If the Agent Registry and Discovery Agent itself is implemented over A2A, even a centralized architecture could support cloud microservices using A2A as the common protocol.
Existing systems such as AWS Agent Registry, MCP Gateway & Registry, and Meta Agent already address parts of this problem, but there is still no open specification comparable to an RFC for DNS.
5. The Important Part Is the Open Specification
We can already build an Agent Registry and Discovery Agent with Codex in a few hours.
The important challenge is not implementation, but defining an open and unified specification for registering and discovering agents, similar to DNS.
Should each of our companies implement its own Agent Registry and Discovery Agent based on such a shared open specification?
6. Reusability Across Environments
With an open and unified specification, we could register and discover agents in a reusable way across different environments.