I'm currently hitting an issue where the agent gets confused (for reasons I'm still investigating) and runs a risk of unnecessarily calling the same tool twice.
The tool in question is not idempotent, and changes application state.
If WebMCP tools had some notion of "turns", one could add some logic that would potentially help tell apart the "Agent got confused and called the tool twice" scenario from "user asked the agent to call the tool twice" one. It wouldn't have been an extremely strong signal, but it'd have been a signal nonetheless.
Is this something other folks would find useful?
Is there a privacy risk here with exposing the number of user interactions with the agent?
I'm currently hitting an issue where the agent gets confused (for reasons I'm still investigating) and runs a risk of unnecessarily calling the same tool twice.
The tool in question is not idempotent, and changes application state.
If WebMCP tools had some notion of "turns", one could add some logic that would potentially help tell apart the "Agent got confused and called the tool twice" scenario from "user asked the agent to call the tool twice" one. It wouldn't have been an extremely strong signal, but it'd have been a signal nonetheless.
Is this something other folks would find useful?
Is there a privacy risk here with exposing the number of user interactions with the agent?