Background
AI agents make it possible to rethink what a data-driven organization looks like. Instead of using AI as a tool inside an existing human workflow, an AI-native organization is built around agents that divide work, generate and analyze data, use computational tools, evaluate one another's outputs, and decide what should happen next. That raises a basic question: how should such an organization operate when it becomes hard to tell whether its work is actually correct?
The team will pick a domain along a spectrum of verifiability: an agentic mathematics institute whose outputs can be formally verified with tools such as Lean, an agentic simulation organization that experiments with computational models, or an incentive-driven organization whose goals can’t be reduced to a simple correctness test. Teams may also propose their own variation.
The project runs inside the SDS AI Studio. At major milestones the faculty mentors act as a review committee, examining demonstrations and results and questioning design choices in studio-style reviews.
Objectives
- Choose a domain where an AI-native organization can pursue objectives through data, computation and agent interaction.
- Design an architecture in which agents take on different roles, use tools and data, evaluate intermediate work, and decide what comes next.
- Define what counts as evidence of success in that domain, and how the organization's outputs can or can't be verified.
- Run experiments on how agents divide work, respond to feedback, recover from mistakes, and decide when they are done.
- Study how organizational design should change as verification becomes weaker, costlier, or more a matter of judgment.
- Deliver a working prototype, reproducible experiments and data, and a final report or research poster.
Milestones
- Early fall 2026: kickoff, choice of domain, research question and measures of success.
- Early to mid fall 2026: studio design review of the proposed organization and first prototype.
- Fall 2026: a working system, demonstration and preliminary findings for the School of Data Science's Datapalooza.
- Spring 2027: research expansion building on fall results and studio feedback.
- Spring 2027: final studio review of complete results and lessons learned.
The fall milestones are deliberately early, so third- and fourth-year students who take part for one semester still come away with a substantial prototype, experiment and presentation.
Who should join
Experience with Python, machine learning, LLMs or agentic AI systems helps. Depending on the domain, formal methods, simulation, reinforcement learning, game theory, data analysis or software engineering may be useful. More important is an interest in how increasingly autonomous AI systems organize work and evaluate their own progress.
- Team size: 3–5 people
- Timeline: 2 semesters, 5–10 hours a week
