Cue quality agents

Agents that show us where Cue helps and where it gets stuck

When Cue launched in 2026, we wanted to learn quickly where it was helping recruiters and where it was getting in the way. So we built a small set of Claude agents. One looks at how Cue is being used and suggests what to fix or try next. Another, which ran from March to July 2026, turned likely bugs into clear, well-documented tickets for engineers. People decide what to act on: the agents never change Cue or contact customers on their own.

01 / 05

The problem

Recruiters and hiring managers rely on Cue to answer questions and take action in Fountain. When something goes wrong, the details sit in logs that take time to read, and feedback buttons alone miss most of those moments. At frontline-hiring scale, no one can review every conversation by hand.

02 / 05

What we built

A learning agent that runs a repeating loop: see how Cue is being used, spot patterns, suggest what to try next, and learn from the team's feedback.
Friction detection that looks for errors and repeated attempts, not just thumbs-down ratings.
A daily briefing for the Cue team during the launch period, with numbered next steps a teammate could start by replying with the number.
An issue agent, which ran from March to July 2026, that sorted problems by type, checked for duplicates so one problem got one ticket, and wrote engineer-ready tickets with what the user was trying to do, what happened instead and possible causes.
A feedback loop that saved engineers' corrections and suggested changes to the agents' own rules. A person reviewed every change before it went in.

03 / 05

How it works

Collect: gather signals about Cue, like errors, retries and feedback, with internal and test accounts filtered out.
Find patterns: spot where Cue delivers value and where people get stuck.
Hand off: likely bugs went to the issue agent, which wrote a clear ticket and sent the most severe to the team that owns that part of Fountain.
Decide: a person chooses which suggestions to act on.
Learn: corrections from the team shape the next run.

04 / 05

Where it landed

Posted regular briefings through Cue's beta and the first weeks after launch in spring 2026.
Filed tickets straight into engineering's existing tracker. Engineering reviewed the first batch of tickets from the rebuilt issue agent and judged the issues valid.
Our customer teams have used the learning agent to answer questions about Cue adoption as they help customers roll it out.

05 / 05

What we learned

Feedback buttons miss most problems. Errors and repeated attempts tell you much more.
Fewer, better tickets beat more tickets. A short list engineers can trust is worth more than a long one.
Say possible cause, not the cause. An agent that sounds certain and is wrong costs engineers time.
Use AI where it adds the most. Simple rules do the sorting, and AI writes up only what is left.
Different jobs need different agents. One spots patterns, another digs into why something broke.
Keep a person in charge of changes to how an agent works, and make it quick to say yes or no.