Skip to main content
The fastest way to improve an Agent is to look at a run and say “no, wrong”. The Trace view is where you do that. Every piece of feedback becomes part of the signal loop that drives Improvements.

​
Adding feedback

Two entry points:
  • The thumbs bar on the final answer (👍 / 👎).
  • Right-click any step in the timeline → Add feedback.
Step-level feedback attaches to that specific step (a wrong tool call, a bad retrieval), not the whole run.

​
Feedback types

  • 👍 / 👎 — quick signal.
  • Text correction — write what the right answer would have been.
  • Categorization — tag with a failure type (hallucination, wrong-tool, stale-info, custom tags).
  • Reference correction — point to the source that has the correct info.
You can combine any of these on a single run. The richer the feedback, the more useful it is downstream.

​
What happens after you submit

The feedback becomes a Signal immediately (see Signals overview). What happens next depends on the type:
  • Standalone signal — a single event, viewable in the Signals queue.
  • Cluster join — if the failure looks like others, joins an existing cluster.
  • Cluster seed — if it’s a new pattern, seeds a cluster of one that may grow.
Rich feedback (with corrections and tags) leads to better clustering and faster improvements.

​
Feedback from users

End users can also give feedback — the Agent’s chat UI has a thumbs bar per turn. User feedback lands in Signals the same way. Anonymous vs. authenticated:
  • Chat sessions from signed-in users carry the user’s identity.
  • Public chat feedback is anonymous by default (only the session ID is retained).

​
Bulk feedback

Two workflows for bulk:
  • CSV import — a spreadsheet with columns trace_id, verdict, tag, correction. Upload from Signals.
  • Programmatic — POST to POST /traces/{id}/feedback. Useful for offline review pipelines or Slack-triggered feedback.

​
Editing / removing feedback

Feedback is editable for 7 days after submission. Beyond that, feedback is immutable — the audit trail matters more than reversibility. If a signal was based on wrong feedback, mark the signal as false alarm (see Confirm or mute) rather than deleting.

​
Feedback and privacy

Feedback shows the trace author’s name (you) unless posted anonymously. Anonymous mode strips identity from downstream views but preserves it in the internal audit log for compliance.

​
Recording feedback as an improvement rule

For team-authored feedback, click Save as rule to promote the correction into a procedural memory rule (see Procedural rules). The rule applies immediately to future runs — the fix is live before Improvements even needs to run. Best used for high-confidence corrections where the fix is easy to state (“always mention ‘before tax’ in price answers”), not for nuanced cases that need proper testing.