Emma: one friendly face, eight agents behind it
The problem
Torre matched candidates and jobs with an algorithm, but reaching candidates, answering their questions and interviewing them still depended on people. A team of about 15 answered 1,200 to 1,500 messages a day with a two-minute response target.
What I did
- Co-designed the agent architecture with an engineer: about eight specialized agents that pass work to each other, behind a single conversation the user sees.
- Wrote the prompts and function calls, and the knowledge base Emma consults, including when she should consult it.
- Placed rule-based checks around the parts where the model was inconsistent, and kept people in the loop until quality held.
Decisions and trade-offs
- One conversation, many agents. A single large prompt answered the same message differently each time. Splitting the work let us test and fix each step on its own.
- People in the loop first. Full automation on day one would have put candidates at risk. For about two months, reviewers checked answers in shifts while we raised quality.
Impact
Emma later grew to run candidate interviews, ask for references, help companies post jobs and learn each user's preferences.
The internal system is private, so this is an illustration of what was built, not the actual product.