Stronger conversation security and more thorough agent evaluations
Your Service agent conversations are now better protected, and agent evaluations are more complete and reliable — giving you greater confidence in every interaction.
What's new in Hub
This update brings two meaningful improvements to Hub's Service agents: stronger access controls for customer conversations, and a more thorough, resumable evaluation process.
Your conversations are better protected
Hub Service agents now apply stricter access controls to every customer conversation.
Previously, knowing a conversation ID was enough to perform actions within it. Going forward, only the person or session that started a conversation can act on it — whether that means approving a suggested action, cancelling a request, or reading a result mid-stream. For signed-in users, this is tied to their identity. For anonymous conversations, it's tied to the specific session that started the chat.
This change means that customers can trust their conversations are private to them, and that agent actions genuinely reflect their intent — not a request from someone else who happened to know a link or ID.
Note for anonymous chat users: Conversations created before this update will need to be restarted. Existing conversation IDs alone will no longer grant access — this is an intentional security improvement.
Evaluations you can actually rely on
When you run an evaluation on a Service agent — checking how well it answers your configured questions — the process is now considerably more robust.
Evaluations now cover all configured questions, running in organised batches and automatically saving progress as each batch completes. If a run is interrupted for any reason, it can pick up where it left off rather than starting from scratch.
The results you see are also clearer:
- Stale results are flagged if the agent's configuration has been edited since the last evaluation, so you always know whether results reflect the current version.
- Partial runs are not treated as complete — a run that didn't finish all questions won't be marked as launch-ready.
- Results now clearly distinguish between source coverage (did the agent find relevant information?) and answer correctness (did it give the right answer?).
These improvements make it much easier to build confidence in your agent before enabling it for customers, and to catch regressions after making changes.
More improvements are in progress. We'll keep you updated as they roll out.