Interleaving user and Agent interaction with the site

[khushalsagar]

Agents doing UI actuation would generally allow only 1 to interact with a site. There's a take over option for the user to take control or the Agent might delegate to the user when something explicitly requires user interaction. We need to think about how this works with WebMCP. Especially since by design WebMCP allows the site to decide when it can execute the tool itself (searchProducts) vs seek user input (makePayment).

[jasonjmcghee]

Should this be combined with #21, or are they more distinct than I'm thinking?

[khushalsagar]

There is definitely overlap. A couple of cases on my mind, let me articulate:

1. Elicitation is the site explicitly passing over control to the user to interact. The site already knows it will be getting user input, the Agent needs to be informed. That's what Elicitation #21 is about.

2. Say the Agent loads the site and is using WebMCP tools to interact with the site. The tools are async so it's possible the user decides to "take over" in the middle of a tool execution. Should the site be informed? What if the user makes a query, the Agent decides it can't fulfill it with the available WebMCP tools and decides to do UI actuation. Does the distinction matter (user takeover vs Agent interacting with the UI).

The main difference I guess is that in 2) the Agent is aware that UI interaction will happen and it's unclear if the site needs to know. I can't think of a concrete use-case for it though so happy to close this issue for now.