Workshop: "AI Agents & the Web"

[ruoxiran]

NOTE: This is a tentative workshop draft — Dom and I are working on a CFP. This document and this issue may be updated, and we would like to hear your opinions in the meantime.

Introduction

This W3C Workshop aims to gather stakeholders to discuss AI Agents and identify potential areas for standardization. The workshop will explore AI Agents in the Web Ecosystem, Inter-Agent Communication, Use Cases and Requirements for AI Agents, and discuss the needs and opportunities for Web standardization to support AI Agents and a possible future Agentic web.

Possible topics:

1. AI Agents in the Web Ecosystem

This topic investigates how AI agents can be integrated into browsers and web applications, and explores the long-term trajectory of AI agents evolving into autonomous user agents or operating systems.

• How should AI agents relate to web browsers and web applications? What level of integration if any needs to be considered between them?

• How might the presence of AI agents reshape user interaction paradigms and information flows on the Web? What risks to the Web ecosystem might AI agents adding a new intermediation layer to content create for publishers?

• How might AI agents evolve into user agents or operating systems, functioning as next-generation browsers or operating systems?

• What APIs (if any) would be needed to support interaction between web applications and AI agents? What inspiration to draw from ongoing development of LLM/browser interaction APIs

2. Inter-Agent Communication Across the Open Web

To enable a distributed and open ecosystem of AI agents, mechanisms for agents to discover, understand, and collaborate with each other may be needed, and the Web would be a natural candidate for that role.

• How can agents reliably identify and locate other agents across services and domains?

• How can collaboration protocols enable agents to exchange intent, task delegation, capability information, and negotiate roles, to enable agents' automated collaboration?

• How should authentication, authorization, access control, and encryption be handled in agent-to-agent communication?

• To the extent that agents act on behalf of end users, what privacy and security mechanisms need to be built into agent communication?

References on this topic:

• AI in the Browser
• Work in W3C AI Agent Protocol CG
• Paper: A Survey of Agent Interoperability Protocols
• Paper: A Survey of AI Agent Protocols

3. Emerging Use Cases and Requirements for AI Agents

With real-world applications of AI Agents already emerging, understanding the value and limitations of current approaches to AI on the Web is critical to manage their impact on the future of the Web.

• How are agents being used today in contexts such as personal productivity, accessibility, Web search, or automation?

• What positive and negative impact are these new agents having on content and service production and delivery on the Web?

References on this topic:

• Work in W3C Autonomous Agents on the Web CG
• AI in the Browser Threat Modelling

4. Standardization Gaps and Opportunities

As emerging new consumers and likely producers of Web content and services, AI agents may raise new demands for interoperability with standardized foundations.

• Where existing Web standards already support the development and deployment of AI agents, are there changes in how we produce and review these specifications to ensure a better fit with that ecosystem?

• How can agent frameworks align with or extend ongoing specification efforts？

• Where do current W3C standards fall short in enabling agentic behaviors, and where are new specifications needed?

• What role (if any) should W3C play in standardization communication and coordination of AI Agents?

Possible outcomes:

• Understand how AI Agents may impact the Web and what role W3C should play in this standardization.
• Explore possible standard of Inter-Agent Communication across domains.
• Identify what updates to existing standards or new specifications are needed.
• Build a roadmap for W3C standardization in the Agentic Web space.

Meeting format: hybrid (tentatively)

[cc'ing @dontcallmedom @plehegar ]

[ruoxiran]

#506

[plehegar]

We need to make sure this workshop and the "Smart Agents" workshop are well differentiated.

[egekorkan]

As further references for topic 3 (possibly also 4), I would like to add that the Eclipse LMOS project, which is about building and running multi-agent systems, is already using 2 W3C RECs from the WoT WG, mentioned at https://eclipse.dev/lmos/docs/lmos_protocol/agent_description/ and https://eclipse.dev/lmos/docs/lmos_protocol/discovery/. Additionally, Agent Network Protocol is using DIDs.

[michaelchampion]

W3C definitely needs to collectively assess where LLM-style AI fits in with the web and the threats/opportunities it creates.
BUT It seems far too late for a traditional W3C academic-style workshop; way too many $ and way too much running code out there, so it will draw more sales pitches than early-stage proposals.
Not sure what to suggest, but some set of Team, AB/TAG, and member experts in a "task force" or whatever you want to call it should:
Define some specific questions
Reach out to member and non-member experts AND IMPLEMENTERS to seek a critical mass of people willing to invest time to help W3C answer them
Come up with answers in a timely and pragmatic manner

[dontcallmedom]

Define some specific questions
Reach out to member and non-member experts AND IMPLEMENTERS to seek a critical mass of people willing to invest time to help W3C answer them
Come up with answers in a timely and pragmatic manner

that's more or less what we're trying to achieve with a workshop, starting with this very issue. W3C Workshops don't have to be "academic-style" - I don't think the workshops I have organized over the years were particularly academic in style (with my limited knowledge of what academic workshops look like)

[michaelchampion]

Whatever it's called -- task force, workshop -- I suggest it should be:

• Much more focused than the current draft.
• Accept that there is no likely W3C consensus on whatever it concludes -- some want W3C to standardize the communications among our new robot overlords, some want W3C to be the Serena Butler of the anti-machine resistance.
• Identify the people you need to have in the room to make it successful and do what it takes to get them there.
• And if you can't get the robot-overlord creators to work at W3C on interop or whatever, maybe you should give up on that and try to get the Serena Butler wannabes to participate

[andreiciortea]

The WebAgents CG has recently launched an Interoperability Task Force whose work should help address some of the points.

We are bootstrapping a living report that we will continuously update as the work develops. We have a rough consensus on the report's scope and structure, and we'll start adding content over the next 2-3 weeks.

The report currently aims to clarify:

• what should be the role of the Web in these developments (at a conceptual level)
• what are the relevant standards within and maybe also outside the W3C
• what are the standardization gaps (if any)

Other notes:

• the scope of the report is broader than LLM agents / Agentic AI, but we have a specific focus on this topic
• the CG has good participation from senior researchers who have been advancing some of the relevant topics
• the CG has good industry participation as well, but we'll need to attract more representation still

We are happy to receive your feedback on how to make this effort most useful.

[jyasskin]

Another question a workshop in this space needs to answer is how we can ensure that any AI Agents support the Web. If they only extract value, people will either stop writing websites, or they'll build Markov pits for the Agents to fall into. Note that money isn't the only support that worthwhile websites need.

[plehegar]

Should we add a list of "potential hazard due to use the AI on the Web" to help encourage submissions of problem statements?

[dontcallmedom]

I've added

What risks to the Web ecosystem might AI Agents disintermediating access to content create for publishers?

in the impact on information flows to reflect @jyasskin and @plh comments

[chrisn]

A note on terminology. Generative AI and AI agents act as a new intermediating layer between publishers and end users, they don't disintermediate.

[dontcallmedom]

indeed, I've amended the text accordingly

[chgaowei]

I am the initiator of the ANP open-source community. ANP (Agent Network Protocol) is an open-source protocol designed for intelligent agents, enabling communication between them. Our goal is to build an open, secure, and efficient Agentic web.
ANP solves identity, description, discovery, and other challenges in agent communication, leveraging W3C specifications such as DID, JSON-LD, and VC.
We can add a topic for Topic 2.
Additionally, several companies are already experimenting with ANP in their products to address agent collaboration challenges.

ANP github URL: https://github.com/agent-network-protocol/AgentNetworkProtocol

[chgaowei]

As further references for topic 3 (possibly also 4), I would like to add that the Eclipse LMOS project, which is about building and running multi-agent systems, is already using 2 W3C RECs from the WoT WG, mentioned at https://eclipse.dev/lmos/docs/lmos_protocol/agent_description/ and https://eclipse.dev/lmos/docs/lmos_protocol/discovery/. Additionally, Agent Network Protocol is using DIDs.

Yes, the Agent Network Protocol uses DIDs. We have studied most identity solutions, including OAuth, API keys, and blockchain, and found that DIDs are the most suitable identity authentication method for intelligent agents. We also use VCs and JSON-LD.

[Thomas-lili]

As further references for topic 3 (possibly also 4), I would like to add that the Eclipse LMOS project, which is about building and running multi-agent systems, is already using 2 W3C RECs from the WoT WG, mentioned at https://eclipse.dev/lmos/docs/lmos_protocol/agent_description/ and https://eclipse.dev/lmos/docs/lmos_protocol/discovery/. Additionally, Agent Network Protocol is using DIDs.

Yes, the Agent Network Protocol uses DIDs. We have studied most identity solutions, including OAuth, API keys, and blockchain, and found that DIDs are the most suitable identity authentication method for intelligent agents. We also use VCs and JSON-LD.

Dear Gaowei, could share why you believe DID/VC is the most suitable identity solution for AI agents? This choice is very important. Thomas

[chgaowei]

As further references for topic 3 (possibly also 4), I would like to add that the Eclipse LMOS project, which is about building and running multi-agent systems, is already using 2 W3C RECs from the WoT WG, mentioned at https://eclipse.dev/lmos/docs/lmos_protocol/agent_description/ and https://eclipse.dev/lmos/docs/lmos_protocol/discovery/. Additionally, Agent Network Protocol is using DIDs.

Yes, the Agent Network Protocol uses DIDs. We have studied most identity solutions, including OAuth, API keys, and blockchain, and found that DIDs are the most suitable identity authentication method for intelligent agents. We also use VCs and JSON-LD.

Dear Gaowei, could share why you believe DID/VC is the most suitable identity solution for AI agents? This choice is very important. Thomas

I previously wrote a blog post comparing the pros and cons of DID, OpenID Connect, and API Keys in the context of agent interactions.
Blockchain wallets are also a potential solution, but currently, it's difficult for blockchain to support large-scale user adoption—such as billions of users.
Here is the article:
https://github.com/agent-network-protocol/AgentNetworkProtocol/blob/main/blogs/comparison-of-did-wba-with-openid-connect-and-api-keys.md

[Thomas-lili]

As further references for topic 3 (possibly also 4), I would like to add that the Eclipse LMOS project, which is about building and running multi-agent systems, is already using 2 W3C RECs from the WoT WG, mentioned at https://eclipse.dev/lmos/docs/lmos_protocol/agent_description/ and https://eclipse.dev/lmos/docs/lmos_protocol/discovery/. Additionally, Agent Network Protocol is using DIDs.

Yes, the Agent Network Protocol uses DIDs. We have studied most identity solutions, including OAuth, API keys, and blockchain, and found that DIDs are the most suitable identity authentication method for intelligent agents. We also use VCs and JSON-LD.

Dear Gaowei, could share why you believe DID/VC is the most suitable identity solution for AI agents? This choice is very important. Thomas

I previously wrote a blog post comparing the pros and cons of DID, OpenID Connect, and API Keys in the context of agent interactions. Blockchain wallets are also a potential solution, but currently, it's difficult for blockchain to support large-scale user adoption—such as billions of users. Here is the article: https://github.com/agent-network-protocol/AgentNetworkProtocol/blob/main/blogs/comparison-of-did-wba-with-openid-connect-and-api-keys.md

Thanks! I agree with your analysis. It's expected that there going to have billions of AI agents over Internet and even more. Under this scale level, simplicity and security are very important and need to be balanced well. The DID/VC mode would be the best way till now. But we need much wider commercailization. W3C needs to speedup the related standardization.
Now 3GPP is discussing 6G's new architecture. Agentic AI Agent + DID/VC + (a new set of protocols among Agents and between Agents and tools/resources) is emerging. This new architecture, if adopted by 3GPP finally, will not only apply on mobile networks, but also mobile terminals and applications. If ANP become part of 3GPP 6G specifications, the adoption will be world wide.

[Wendong-Fan]

Dear W3C Workshop Organizers,

I'm a core contributor and Tech Lead for CAMEL-AI. We are very interested in contributing to Topic 1: AI Agents in the Web Ecosystem.

CAMEL-AI (https://www.camel-ai.org/) is an open-source research community dedicated to building foundational infrastructure for AI agents. We focus on collaborative, tool-using agents designed to operate in complex environments, including interacting with web browsers, APIs, and other web-based tools. Our work aims to empower researchers and developers to rapidly build, experiment with, and understand multi-agent systems.

Looking forward to further discussion!

[anolan4]

Thanks for organizing this Ruoxi and Dom. I worry the current text is not scoped enough to lead to productive conversation in a workshop. It might be helpful to include a tangible outcome of the workshop to focus more on progress than academic-style conversation. One idea would be that perhaps the workshop could result in a whitepaper focused on various sections like landscape, benefits, risks, challenges, and open questions related to AI LLMs and agents interacting with the web. The focus would be on the web, not a single company or technology. The whitepaper would be reviewed by various W3 groups including AB and TAG for eventual publishing publicly to clearly provide a web perspective on this important topic.

[michaelchampion]

@plehegar wrote:

Should we add a list of "potential hazard due to the AI on the Web" to help encourage submissions of problem statements?

I'd suggest focusing on the hazards that actual LLM AI agents are already creating. You don't need a lot of "problem statements" you need to identify concrete problems that the W3C community can collectively address.

I'd also suggest trying to discourage "solutions looking for a problem" submissions from those hoping to get market credibility from a W3C standardization effort. This field is moving MUCH FASTER than W3C has ever managed to operate; as I understand it (disclaimer: as a retired, unaffiliated, non-expert) Anthropic's MCP has become the de-facto standard LLM agent API less than six months after its introduction. Even more than during the current web platform's heyday, industry standards came from INDUSTRIES, with standards organizations coming later along to clarify / test / certify. The primary goal of a workshop MUST be to attract the people doing actual innovating and deployment in whatever focus area you choose, and build a community of people who want to find common ground to address common concerns.

@anolan4 wrote:

I worry the current text is not scoped enough to lead to productive conversation in a workshop.

+1000. I agree with @dontcallmedom that he has organized some successful workshops, but they were FOCUSED on CONCRETE problems, and got REAL EXPERTS together to share thoughts on how to address them. The current draft has more of the flavor of much earlier W3C "academic-style" workshops that shared diverse perspectives but didn't end with clear consensus on how to tackle actual problems.