Clip
Summary
The team is building user-delegated consent into connectors. Whenever an agent tries to invoke a connector tool, the user is prompted — approve once, approve all for that connector, or deny. You pick the strictness: no prompts, write operations only, or every call. Alex showed the agent debugger pausing mid-run waiting for permission. Connectors will need to adopt the new API, but the responsible-AI pattern is taking shape.
Screenshots
Transcript
that agents are running differently between the assistant builder chat interface and the quick asks. And this will essentially solve that problem because we're essentially running the same back end is going to be used. Across both systems, and then we can obviously extend to other systems running these agents and build experiences in a similar way.
Now for my second demo, we've also worked on... building experiences around user consent. I'm going to quickly refresh here. And for this, essentially, we've looked, we've modified connectors to intercept the tool invocation in the context of a user from an agent.
And look if the user has previously consented to this to this operation or not. Now, there are some sort of API changes whenever we use this new mechanism in connectors, so we will need to work with the teams integrating directly with connectors to adopt this new mechanism, but I'm going to show what's the... Experience in this in this new chat as well, so I'm going to select the agent debugger.
So, again, this only works with agents that are using connectors, so not going directly to the to the to their MCPs or open API. plugins, but I think, yeah, the connector which is coming up, I think, will help us model a lot of those products as connectors as well. And then here, the API has a few options. How do I want to consent?
I want no consent prompts, I want write operations to be to require consent, or I want all operations to require consent, which is what I'm going to select here, and I can ask my agent debugger to show me the list of agents. And this will go ahead and use the connector to retrieve that information. And I can see here that I am being asked for an approval.
The agent needs my permission to execute this operation on this specific connector. And here I have options to deny this operation. to approve for this specific operation or to approve all the operations for this specific connector. So in this case, I'm going to click approve.
And after I click approve, the agent will continue on what it was doing and essentially. continue the workflow from here. Now, in the next iteration here, it might need to call a different operation and that would have a separate. consent. So all the consent are per operation basis at this moment and each would need to be consented individually.
So this is a standard best practice for things like human in the loop and yeah. sort of responsible AI principles.