Outcomes & feedback
An agent turn that completed without an error is not the same as an agent turn that helped. Yekar.AI records the difference in two places, and both feed the Health page and Learning.
How the conversation ended
The session's creator/recorded owner or a domain Owner can set or clear its outcome in the details panel. Other viewers can inspect the outcome:
| Outcome | Use it when |
|---|---|
| Resolved | The person got what they came for. |
| Unresolved | They did not, and nobody picked it up. |
| Escalated | A human took it over. |
A session nobody has judged stays unset, and that is a distinct state - not the same as unresolved. Every read that reports on outcomes states its coverage first: what fraction of sessions anyone has judged at all. A resolution rate over a handful of judged sessions and one over thousands look identical as percentages and mean nothing alike, so the count travels with the rate.
This is the primary outcome signal because it works for every session, including the ones nobody watched.
Rating a reply
Beside any agent reply, mark it Good answer or Needs work.
A thumbs-down asks what went wrong, from a fixed vocabulary:
- Incorrect - the answer was wrong.
- Incomplete - right as far as it went.
- Not supported by its sources - the citations do not carry the claim.
- Wrong tool or arguments - it acted, but not the way it should have.
- Refused something it should do - it declined work it was configured for.
- Something else.
The vocabulary is fixed on purpose. A closed set is countable - you can ask "how often is this agent wrong about its sources" and get an answer - and it keeps the record to facts rather than free prose, which is what lets these rows sit alongside the rest of the audit trail under the same handling rules.
Ratings describe the interactive minority. Most turns run unattended, which is why the session outcome above is the primary signal and a thumb is the secondary one.
What the signals drive
- Health reports where answers came from and what was repaired, cut by Setup revision.
- Learning collects the runs that went badly into a reading list, and proposes rules from them.
- Knowledge gaps are built from turns people were happy with whose answers did not come from your library - the one place "the knowledge base should have covered this" is demonstrably true rather than a guess.
A workspace that never marks outcomes still gets working agents. It just cannot tell you whether they are getting better.