A structured operating context for working with AI agents in doctoral research.
AI agents now contribute to doctoral work, but mostly under ad hoc arrangements. The research harness specifies the agent’s operating context in seven components, so the researcher stays the analyst, the judge, and the author of their own inquiry.
What it is
AI agents now contribute to doctoral work — but mostly through individual exchanges governed by no shared account of what the agent is doing across the project as a whole. The familiar troubles follow from that absence: work running ahead of the thinking it depends on, the quiet offloading of interpretation, a project drifting as each plausible suggestion accumulates, and, later, no way to say who contributed what. Institutional policy and AI-literacy training don’t reach this, because the gap sits between the institutional rule and the individual exchange.
The research harness borrows an idea from software engineering, where agents working without a defined operating context produced the same kinds of trouble. It is a structured specification of that context, in seven components: a knowledge base, interpretive permissions, tools, authority, a scope register, a process record, and an amendment protocol. You can start thin — a single sentence under each — and grow it as the work demands.
What it does
The harness is a developmental scaffold, not a control mechanism. Throughout, the researcher remains the agent’s analyst, its judge, and its author; the agent contributes to the inquiry without taking it over. Because it is a shared artefact that supervisor and researcher negotiate together, it gives supervision something concrete to work with — a way to make the terms of AI use explicit, reviewable, and revisable as a normal part of doctoral development. The process record adds accountability and provenance: a versioned log of what was asked, produced, and decided, and why.
The underlying commitment is that a doctorate develops the researcher, not just the thesis — and the skills needed to evaluate AI’s output are the very ones a research training exists to build. The harness is designed to protect that development rather than let it be quietly delegated.
Status
The framework is published as a preprint on the Open Science Framework and condensed into a one-page guide for researchers and supervisors. It is now being developed into a tested intervention: refining the framework toward a v1.0 freeze, building an open-source agent that walks a researcher through setting up a harness, and designing field tests of whether specifying the harness actually changes research practice — not just whether it sounds like it should.
Find out more
- Essay: The research harness: a framework for bounded AI use in doctoral work
- Guide: The research harness — a one-page guide for doctoral researchers (downloadable PDF)
- Preprint: Open Science Framework