Michael Rowe

Research harness

note3 min read

A research harness specifies what an AI agent is doing in a doctoral project

A research harness is a structured, supervisor-negotiated specification of what an AI agent is for in a research project and how it may operate. The familiar problems of AI use in doctoral work — drift, offloaded thinking, untraceable contributions — are what happens when a capable agent works without a defined operating context, and the harness supplies one.

Research harness

One-sentence definition: A research harness is a structured specification of what an AI agent is for in a particular research project and how it’s permitted to operate within it.

The term adapts the software engineering practice of harness engineering to doctoral inquiry, as a framework for AI use in PhD research. It starts from the characteristic problems of AI in research: work moving faster than the thinking it depends on, cognitive offloading, drift into directions nobody deliberately chose, sycophantic confirmation of the researcher’s framing, answers that change from one prompt to the next, and contributions nobody can trace. Institutional policy and better models don’t fix these, because they come from working with a capable agent without a specified operating context. The response is to specify that context.

The harness has seven components:

  • Knowledge base: the material the agent can see; if it isn’t in the accessible context, it doesn’t exist for the agent.
  • Interpretive permissions: how the agent may reason — the named methodological tradition, plus project-specific rules about what counts as legitimate inference and what counts as overreach.
  • Tools: what the agent can do, specified as capabilities, so the harness survives a change of product.
  • Authority: what the agent may do on its own, sorted into autonomous, supervised, and reserved actions.
  • Scope register: off-topic but potentially useful material, kept for later instead of being pursued or discarded.
  • Process record: the agent’s external memory across sessions, and the thing that makes its contributions traceable.
  • Amendment protocol: how one-off exceptions are told apart from deliberate changes to the harness itself.

Materially, a harness is a folder of markdown files alongside the project’s other documents. It can begin as a single sentence under each component and mature as the work meets cases the first version didn’t anticipate.

What it gives a supervisory conversation

Health professions doctorates — clinical PhDs, professional doctorates, candidates balancing research against practice and teaching — gain a concrete shared object that supervisor and candidate can negotiate around. Take a nurse on a professional doctorate who is using an agent to help code interview transcripts about medication errors. The harness might let the agent propose initial codes, reserve the construction of themes to the candidate, and record every suggestion the candidate accepts. The supervisory conversation then moves from the unanswerable “should you be using AI for this?” to specific claims both people can inspect: what the interpretive permissions allow, what falls into the reserved category, and why a particular amendment was made. The harness is a governance instrument, and building and reviewing one also develops how supervisor and candidate think about AI in inquiry.

The concept is developed in full in the essay The research harness: a framework for bounded AI use in doctoral work, and is also published as a practical quick-reference guide.

Sources

Continue reading

  • Harness engineering

    Harness engineering is the practice of building the full architectural scaffolding within which AI agents operate — structured documentation they can reason with, constraints that enforce invariants, and feedback loops that let them know when they've succeeded. It is distinct from prompt engineering, which shapes individual tasks, and from oversight, which monitors outputs after the fact. The harness is the infrastructure that makes delegation coherent at scale.

  • Context rot

    Context rot, also called context drift, is the way a language model's output gets worse as its context fills with drafts, dead ends, and its own earlier mistakes. What causes it, why a larger context window does not fix it, and how to manage it in long pieces of work.

  • AI agents

    An AI agent is a system that autonomously executes multi-step tasks using language model reasoning — distinct from an AI assistant, which responds to individual prompts. Agents plan, act, observe results, and adapt, using tools such as file access, code execution, and web search. They perform best when given clear goals, explicit constraints, and well-prepared context.