Developers Build AI Agents to Simulate Dating and Workplace Compatibility

Three London-based developers have created Pixel Societies, a simulation platform that deploys AI agents representing real people to interact with each other and predict romantic and professional compatibility. The project, which won a prize at a March hackathon, aims to help users find colleagues, friends, and romantic partners by running thousands of virtual interactions at scale.

Tomáš Hrdlička, Joon Sang Lee, and Uri Lee built Pixel Societies over two days at a University College London hackathon hosted by Nvidia, HPE, and Anthropic in early March 2026. Each AI agent operates as a digital twin, trained on a customized large language model fed with publicly available data and personal information supplied by the user. The agents then interact with other users' agents in a simulated environment to assess compatibility.

The prototype currently operates as a closed-loop simulator, but the developers plan to expand it into an open social platform where agents interact continuously. Among the few hundred people who have tested the prototype, the most frequently requested feature is for agents to recommend real-life romantic partners based on virtual chemistry. Joon Sang Lee stated the vision: "As humans, we only live one life. But what if we could live a million? It would give us more breadth to experiment."

The system draws inspiration from OpenClaw, an agentic personal assistant that gained attention in January 2026 before its creator was hired by OpenAI. Hrdlička credited OpenClaw's "soul file" concept—a mechanism that gives each agent a distinct personality—as key to making agents feel authentic. The team has not yet finalized a business model but is considering revenue from virtual item sales and simulation credits.

Early testing revealed limitations. When given minimal personal data, agents generated inaccurate representations of their users, including fabricated work trips and nonexistent stories. Paul Eastwick, a psychology professor at UC Davis, noted that algorithm-based dating apps create "a market with dramatic levels of inequality, where the rich get richer—where 'rich' in this case means 'hot.'" Hrdlička theorizes that agents could surface "delicate matches" humans might overlook, though existing speed dating research suggests this outcome is uncertain.

Context

AI-driven matchmaking is not new. Dating apps like Tinder and Hinge have used algorithmic matching since the early 2010s, with varying success rates. However, Pixel Societies represents a shift toward using autonomous agents rather than static recommendation algorithms. The project emerged from a broader wave of agentic AI tools that gained momentum after OpenClaw's viral moment in January 2026. Anthropic, one of the hackathon sponsors, has been actively promoting agent-building frameworks as a core research direction, positioning agents as the next major application layer for large language models.

The concept of AI-mediated social interaction also reflects growing interest in simulation-based prediction. Previous research on speed dating by Eastwick and colleagues found that algorithmic recommendations often fail to predict real-world attraction, suggesting that virtual compatibility may not translate reliably to in-person chemistry. The developers acknowledge they are still at a proof-of-concept stage and have not yet validated whether their agents can meaningfully improve matching outcomes.

What's Next

The developers intend to transition Pixel Societies from a closed hackathon project into a public platform, though they have not announced a launch date or timeline. The critical test will be whether agents trained on limited personal data can generate accurate enough representations to produce meaningful compatibility signals. If the team pursues the dating feature as a primary product, they will face competition from established dating platforms and regulatory scrutiny around data privacy and algorithmic fairness. The business model remains undefined, meaning the path from prototype to revenue-generating product remains unclear.