xTown: generative agents for recruiting

Inspiration

Traditional hiring uses 4-5 interviewers making a decision with sparse data. But what if a huge number of AI agents—each embodying a real employee—could evaluate every candidate, form independent opinions, and continuously learn from outcomes?

Stanford's Generative Agents paper showed AI can simulate believable human behavior through memory, reflection, and planning. This project attempts to apply this to the recruiting problem in business.

What it does

xTown is a multi-agent simulation environment where candidates are "dropped into" a virtual company to interact with AI-powered employee personas.

Core loop:

Candidate agent (built from LinkedIn/X/GitHub) enters the simulation
employee agents (built from real xAI employee data) interact naturally with them
Recruiter agent conducts structured interviews
Each agent forms independent assessments using memory-augmented reasoning
System aggregates opinions → ensemble recommendation
Human feedback triggers online policy updates

Output: Team-matching recommendations with explainable reasoning from multiple perspectives—not a black-box score.

How we built it

Generative Agent Architecture:

Memory Stream: Timestamped observations with recency/importance/relevance scoring
Reflection: Periodic synthesis triggered when accumulated importance > threshold
Planning: Hierarchical action decomposition for goal-directed behavior

Challenges we ran into

Due to data limitation, we can't collect actual persona from employees and candidates. The data used are from the slack channel #team-matching's self-introductions and public data.