A great day at the WWW conference where I had the chance to present at the special event on 𝐑𝐞𝐬𝐩𝐨𝐧𝐬𝐢𝐛𝐥𝐞 𝐚𝐧𝐝 𝐒𝐚𝐟𝐞 𝐖𝐞𝐛-𝐛𝐚𝐬𝐞𝐝 𝐀𝐈 𝐀𝐠𝐞𝐧𝐭𝐬.

A central, and perhaps contrarian, view I shared: 𝐭𝐡𝐞 𝐟𝐮𝐭𝐮𝐫𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐖𝐞𝐛 𝐢𝐬 𝐟𝐨𝐫 𝐀𝐈, 𝐧𝐨𝐭 𝐡𝐮𝐦𝐚𝐧𝐬. Increasingly, humans will converse with agents, and those agents—not humans—will interact with the Web. We are already seeing websites optimising for agents rather than people, and not just for efficiency. Manipulation risks—akin to SEO for LLMs and agent-targeting deception—are rising fast.

The agent AI community is also, often unknowingly, reinventing the Web’s long history: from stateful AI protocols to agent-to-agent communication (A2A). But we have a rare opportunity now. Unlike the Web’s early years—where trust and security were bolted on only after harm emerged—this time, we can embed trust by design into AI agents and agent-to-agent systems.

I also highlighted two exciting pieces of work from CSIRO’s Data61 team:
– UI understanding for agent context: Our SeeAction paper (ICSE 2025 Distinguished Paper https://lnkd.in/ghYYNcEr) helps (GUI) agents interpret human behaviour and interface context without invasive integration.
– Dark pattern defence for AI agents: Building on our pioneering detection of manipulative design against humans, we are now extending this to defend agents from adversarial web design.

Finally, while many debate the limitations of web data and the risks of synthetic data, I argued that—as we’ve seen with the rise of “zero” models (from the early AlphaGo Zero to R1-zero – meaning zero human data/labelling)—models relied heavily on synthetic data are improving fast. We should not underestimate the power of synthetic data to fuel smarter agents.

Slides https://www.dropbox.com/scl/fi/r98qd117jy9i2vi05ugjf/20250501-Web-based-Agents.pdf?rlkey=hkc8ny6hm4kbrlnpkl6j8cdb2&dl=0