Research

My research concerns AI agents, both as models of human decisions and behavior and as systems that act on their own or alongside people. I am interested in how they use information to make decisions and how the consequences shape their next choices.

So far, my research has focused on the static version of this problem. With Abel Iyasele, I compared responses from three open-weight models (Gemma 3, Bielik and Qwen 3) with World Values Survey Wave 7 distributions across 63 demographic personas and 31,500 stored responses. We measured the gap using normalized Wasserstein distance and bootstrap confidence intervals, then tested how LoRA fine-tuning changed the fit across groups. It improved every targeted persona for Bielik and worsened four of five for Qwen. For Bielik, helping the targeted groups also came with worse representation of some other groups. An extended version of our OSSConf 2026 paper is available as an arXiv preprint.

This work grew out of my undergraduate thesis at SGH Warsaw School of Economics, where I compared Google Gemma 2 with WVS responses across 10,692 demographic subgroups. In a separate election project, I worked with a panel of 250 counties across Pennsylvania, Michigan and North Carolina. Testing on a held-out state exposed poor generalization, sharpening my interest in what a model captures about its setting and what carries over when that setting changes.

I'm always looking for a real-world use for the systems I build. Giving a model four or five traits is a useful starting point, but I find it hard to treat that persona as a person. The survey work showed me how much room there is to be creative in how we engage with models. It also made me want to move beyond a question and an answer.

What people say is only one part of how they relate to each other. How people act and respond to each other matters too. I want to bring more of that into the systems I build, taking inspiration from situated cognition to create agents grounded in human data whose actions have consequences. I think we now have the tools to build much richer simulations and test what they capture about people. That is a big part of why I want to do a PhD.

I want to develop this work in three connected directions.

Agent decision-making

How do memory, beliefs, attention and planning shape an agent's decisions? I want to investigate how agents make decisions with incomplete information and revise them as they learn more. Unfamiliar situations are especially interesting: can an agent recognize that its plan no longer works and use that experience to change course?

Where an agent represents a person, I want to ground it in what that person reports about themselves or in records of their behavior, then test its predictions on decisions it has not seen. There is already evidence for going beyond demographic labels (Park et al., 2026). One starting experiment would compare agents built from different sources of information about the same participants under a matched information budget. I would like to establish what that extra information contributes, using existing human data or working with collaborators where possible.

I am also interested in evaluation through simulations and decision games with unavoidable tradeoffs. Models can pick up cues that they are being evaluated (Needham et al., 2025), so I want to test how choices change with the setting as well as with the task. How does an agent choose between competing goals when resources are limited? At what point does it ask for help?

Dynamic simulations and interventions

I want to build simulations in which the consequences of an agent's actions become part of its next decision. How does behavior change when the rules of an environment change? Even under the same rules, different access to information or resources could affect individual choices and collective outcomes. I am equally interested in changing the agents themselves: their memory, goals, planning horizon or decision rules. I would like to compare language-model agents with agents trained through reinforcement learning, using simpler behavioral rules as a baseline.

For example, people leaving a large event might be heading to different destinations when a route suddenly becomes unavailable. With limited information about the disruption, how do they replan? I want to compare how different ways of replanning change the resulting congestion, and whether an intervention still helps when the assumptions change. Cities and mobility are one setting for this; the broader question is how adaptation works when people or agents share constraints and resources.

Does everyone pursuing their own interests lead to a better outcome for society, or can it leave us all worse off (Kollock, 1998)? I want to use agent simulations to study when individual and collective interests align and how interventions change that relationship (Vezhnevets et al., 2023).

Coordination and human–agent collaboration

I am also interested in how agents divide work and keep each other informed, especially when something goes wrong. What happens when their goals conflict or they remember the same interaction differently? I want to study how access to tools affects their ability to coordinate, both with other agents and with people. Coordination can happen through communication, but also through actions and their effects on a shared environment.

My experience building an analytics MCP server and supervising coding-agent workflows gives me a practical foundation for this work. At Faex, I used written specifications to guide work and handoffs between coding agents, keeping a record of 142 decisions. I directed and reviewed work by Claude Code, Codex and Kiro. I want to study how assigning tasks and tools affects error detection and recovery, and where human oversight makes a difference. The suggestion that an LLM may use “a different background concept” from human annotators reinforces my interest in how people and agents interpret shared instructions (Schroeder, Roy and Kabbara, 2025).

Across these directions, I want to understand when simulations can predict behavior and when they are more useful for developing explanations worth testing. I see room for exploratory work alongside stricter tests of behavioral prediction (Anthis et al., 2025). When agents represent people, I want to check how closely their choices match real human decisions. When agents carry out tasks, I want to see how well they work together and where they fail. In both cases, I want to know how much we can trust the results and whether they hold in new situations.

I’m currently working on Agent reliability under disruption, studying whether agents representing real people can adapt when their daily routines change.