Regulating with AI Agents: A Map and an Agenda

Michaelmas 2026 project

AI capabilities are advancing rapidly and unpredictably. This progress is bringing about significant risk, ranging from enabling malicious actors to conduct sophisticated cyberattacks to misaligned systems pursuing goals their developers did not intend.

That AI needs to be governed is uncontroversial, yet how to do so is contested. There is little consensus regarding which mechanisms will prove both effective and legitimate at governing these technologies. This is due to the increasing technical complexity of systems, the scale at which they are deployed, and the difficulties regulators face in monitoring them.

One potential route for addressing these challenges is using AI agents to support regulation. Frontier labs already deploy AI agents to monitor and audit their systems and how they are used. Public bodies are also increasingly using AI for their regulatory and supervisory functions. Given these trends, it is reasonable to anticipate that the use of agents in support of regulation will grow. Yet, relying on agents for governance raises myriad practical and ethical issues, including concerns about reliability and regulators’ dependence on the companies they are meant to oversee.

This project critically examines the opportunities and risks of using AI agents in support of regulation. It will focus on:

  • Mapping the different governance functions that agents could perform (e.g., behaviour monitoring, rule operationalisation, incident aggregation);

  • Considering how these functions could be operationalised, and by whom;

  • Evaluating the practical and ethical issues associated with these approaches;

  • Providing an agenda for future research on regulating with AI agents.

Undertaking this research is important because of the risks AI poses, the need for regulatory oversight to address them, and the continued uncertainty over how to practically achieve this. Utilising agents could strengthen oversight, but may also produce unintended harms. Systematic research is thus needed to understand whether, and under what conditions, AI agents can support effective and legitimate governance.

Huw Roberts