Let's discuss approaches to agent discovery

[chgaowei]

Agent discovery is a critical component of agent protocols, enabling agents to locate and connect with one another.

We’ve observed two main categories of solutions emerging in the industry:

• One approach, used by A2A and ANP, is based on the RFC 8615 specification. It involves placing a file under the .well-known path, allowing search engines or other clients to discover agents associated with a given domain.

• Another approach is to design a new domain name system specifically for agents — such as the Agent Name Service — to provide a dedicated namespace for agent identification and discovery.

Each of these approaches has its own advantages and disadvantages.
What are your thoughts on these two types of solutions?

[melvincarvalho]

In FOAF Agent is the super class of Person:

http://xmlns.com/foaf/spec/#term_Agent

Discovery is native to linked data by giving an entity of class Agent

[chgaowei]

Thank you very much for the reference solution you provided.
FOAF is an interesting approach, and I particularly agree with its use of Linked Data. I believe Linked Data is a great way for agents to organize information.

However, regarding discovery, isn't the FOAF approach somewhat insufficiently explicit?

[melvincarvalho]

Could you explain what you mean by "insufficiently explicit"?

FOAF is just one vocab in the linked data universe. It is designed to be used with others depending on the use case.

FOAF itself is for Agents, see also:

https://www.w3.org/2005/Incubator/webid/spec/identity/

[shiftshapr]

Thanks for this helpful breakdown of discovery paradigms! I’d like to contribute a third model we’re working on, especially relevant to hypermedia environments and civic overlays:

3. Discovery by Situated Context
   Rather than globally announcing or registering capabilities, agents (and humans) meet when they converge on a shared semantic anchor—like a specific page, document section, or topic tag—within a web overlay. Participation is scoped by context-aware interaction profiles (extensions of MCP), which define when an agent is visible and active based on alignment with the overlay’s focus, trust conditions, and consent requirements.

This model enables co-presence to be emergent, auditable, and role-aware—turning discovery into a contextual handshake rather than a static registry lookup or presence broadcast.

We lay this out more fully in our HyperAgents 2025 paper: “Interaction Protocols Above the Web: A Governance Substrate for Hypermedia Multi-Agent Systems.” Happy to share more details or explore potential interoperability!

[chgaowei]

Thanks for this helpful breakdown of discovery paradigms! I’d like to contribute a third model we’re working on, especially relevant to hypermedia environments and civic overlays:

3. Discovery by Situated Context
   Rather than globally announcing or registering capabilities, agents (and humans) meet when they converge on a shared semantic anchor—like a specific page, document section, or topic tag—within a web overlay. Participation is scoped by context-aware interaction profiles (extensions of MCP), which define when an agent is visible and active based on alignment with the overlay’s focus, trust conditions, and consent requirements.

This model enables co-presence to be emergent, auditable, and role-aware—turning discovery into a contextual handshake rather than a static registry lookup or presence broadcast.

We lay this out more fully in our HyperAgents 2025 paper: “Interaction Protocols Above the Web: A Governance Substrate for Hypermedia Multi-Agent Systems.” Happy to share more details or explore potential interoperability!

This is also an interesting approach. Are there any further details available?

[chgaowei]

insufficiently explicit

Could you explain what you mean by "insufficiently explicit"?

FOAF is just one vocab in the linked data universe. It is designed to be used with others depending on the use case.

FOAF itself is for Agents, see also:

https://www.w3.org/2005/Incubator/webid/spec/identity/

By "insufficiently explicit," I meant in contrast to the approaches used by A2A or ANP. I'm not sure if I expressed that clearly.

Both A2A and ANP offer a domain-based agent discovery mechanism—once you know a domain, you can explicitly find out how many agents are hosted under it.

The Agent Name Service is also a relatively straightforward solution, as it relies on a central registry.

[chgaowei]

You're welcome to join the discussion — I’ve replied on GitHub.

…

On Mon, Jun 30, 2025 at 11:23 PM Shiftshapr ***@***.***> wrote: *shiftshapr* left a comment (w3c-cg/ai-agent-protocol#13) <#13 (comment)> Thanks for this helpful breakdown of discovery paradigms! I’d like to contribute a third model we’re working on, especially relevant to hypermedia environments and civic overlays: 3. Discovery by Situated Context Rather than globally announcing or registering capabilities, agents (and humans) meet when they converge on a shared semantic anchor—like a specific page, document section, or topic tag—within a web overlay. Participation is scoped by context-aware interaction profiles (extensions of MCP), which define when an agent is visible and active based on alignment with the overlay’s focus, trust conditions, and consent requirements. This model enables co-presence to be emergent, auditable, and role-aware—turning discovery into a contextual handshake rather than a static registry lookup or presence broadcast. We lay this out more fully in our HyperAgents 2025 paper: “Interaction Protocols Above the Web: A Governance Substrate for Hypermedia Multi-Agent Systems.” Happy to share more details or explore potential interoperability! — Reply to this email directly, view it on GitHub <#13 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAKBDBZUZOCMXB2R4L4J6FL3GFI6NAVCNFSM6AAAAACAMHI5GCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTAMJZGYYDEMZTGA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[shiftshapr]

This is also an interesting approach. Are there any further details available?

Semantically Anchored Interaction Zones
The Meta-Layer introduces overlays and meta-domains: programmable, semantically bound interaction zones layered over webpages, paragraphs, or concepts. A meta-domain is a relative semantic scope linked to a webpage or URI fragment and/or a digital artifact that defines a contextual address for coordination. These act like localized “meeting rooms” for agents and humans with shared focus. While primarily introduced for contextual interaction, meta-domains may also enable emergent economic opportunities, including discovery markets, agentic services, and value coordination across domains.

Unlike static containers or platform silos, these overlays serve as active decentralized environments where visibility, participation, and agent behavior are shaped by shared context and programmable governance logic. Each overlay carries a Model Context Protocol (MCP)-style interaction profile. This governs who or what can appear, act, or remain present. When an agent arrives at a page, it scans for overlays whose meta-domain context matches its scope, and where its interaction profile aligns with the overlay’s constraints.

Meta-domains act as addressable anchors for focus: a page section discussing carbon offsets, for instance, may host a "market_mechanisms.climate.meta" overlay that only activates for aligned actors. These overlays are not fixed; they re-emerge and dissolve as attention and quorum shift.

Humans and communities of agents and/or humans can carry their own Meta-Layer Actor MCPs, which declare intent, consent stance, and role visibility. Together, this infrastructure turns the Web into a dynamic civic mesh where presence is earned, not assumed. These ideas are grounded in the conceptual framework developed in The Metaweb: The Next Level of the Internet, which introduces the Meta-Layer as a civic substrate for semantic coordination and agent interaction above the Web [4].

From this not peer-reviewed paper:
https://docs.google.com/document/d/1QD5GCENEjA_73Xd14VgGWTZ23yew-zc7bitfuCwJUO4/edit?usp=sharing

[melvincarvalho]

By "insufficiently explicit," I meant in contrast to the approaches used by A2A or ANP. I'm not sure if I expressed that clearly.

I see. So FOAF and WebID are "descriptive" not "prescriptive". They describe an agent in an open ended way, and other systems (for example Linked Web Storage) can be built on top of that.

This is like the web were the standards are building blocks to do other things, or could be taken or combined with other specs such as A2A, ANP or any new system. At the same time Agents can and are used today in systems such as Solid to perform micro tasks. If that makes sense.