Support agent to agent interaction

[khushalsagar]

Forking for the comment made here. We should definitely be thinking of how an interaction between the browser agent and a site agent would be coordinated.

Here's how I was naively thinking about using an MCP tool to expose functionality implemented by an Agent.

window.agent.registerTool({
  execute: callProductFinderAgent,
  name: "product_finder_agent",
  description:
      "An Agent which can perform high level tasks. The parameter is a string describing the task. Here's example tasks for this Agent.
       1. Provide shoes which will look good with a floral dress.
       2. Filter products on the current page to what can be delivered by tomorrow.",
  inputSchema: {
      type: "object",
      properties: {
          "query": {
              type: "string",
              description: "the description of the task the Agent needs to perform",
          },
      },
  },
});

The site Agent should be able to take multi-modal input like images for queries like, "give me products which look like this", but that has to be supported for tools in general.

@igrigorik curious on your take here? I haven't dug into the agent-to-agent protocol which will clarify where this falls short but it could at least unblock basic agent to agent coordination.