Part 3: AI Harnesses for Education: How do we Responsibly Implement This Emerging Layer of AI Infrastructure

Authors: Pati Ruiz, Jordan Mroziak, Keun-woo Lee, Megan Pattenhouse

In part 1 of this series, we discussed what an AI harnesses for education is, in part 2, we discussed why we need AI harnesses for education. In this final post of this 3-part series, we will dig deeper into the consideration for an AI harness for education. Most importantly, we want to make it clear that we don’t believe that a harness as the appropriate location for every desirable property of educational AI; but that it shows promise at addressing issues around pedagogical soundness, transparency, and reliability.

Like most emerging educational technologies, AI harnesses for education and how we responsibly develop them will depend on many factors and require co-design with educators. Additionally, harness-level mechanisms introduce their own risks. So, for example, a system that strictly controls when learners can access assistance may preserve particular forms of cognitive effort while reducing learner autonomy. An orchestration system designed around one theory of learning may constrain educators who use a different instructional approach. Pulling in multiple sources of data would allow for more responsive support while creating privacy concerns.

Consequently, the question of what belongs in an educational harness is both technical and pedagogical. Three key considerations for educational harness design:

• First, the educational objective must be sufficiently clear to be represented technically. Broad aspirations such as “individualized learning” or “student-centered instruction” do not by themselves specify system behavior. More precise constructs, such as requiring an initial learner attempt before a particular form of assistance becomes available is more readily implemented and evaluated.
• Second, the mechanism should be evaluated for its effect on learning and instruction, not only for its effect on agent performance. A harness that makes an agent more reliable at completing tasks does not necessarily make the resulting system more educationally effective.
• Third, decisions about control should account for the roles of learners and educators. Educational systems are not simply environments in which an autonomous agent acts. They are social systems in which authority, interpretation, responsibility, and judgment are distributed among people and technologies.

These considerations suggest that educational harness development would benefit from collaboration among AI engineering, learning sciences, educational technology, human-computer interaction, educators, and learners. The technical literature can identify mechanisms through which model behavior can be structured. Educational research is necessary to determine which behaviors are desirable, under which conditions, and how their effects should be evaluated.

The problem of Observability

Observability is a central concern in emerging harness engineering. In technical systems, logs and traces can be used to identify failures, reconstruct agent actions, inspect tool calls, evaluate costs, or determine whether system constraints were followed

Educational applications may require related forms of observability, but the relevant evaluation criteria are different. For example, successful completion may be insufficient to understand the learning process someone has undertaken. A learner may independently reason through a problem after receiving a focused hint, or may obtain a complete solution from an AI system and reproduce it. Evaluating only the final product obscures this distinction.

Harness-level traces could potentially make aspects of these processes available for analysis. Records might include the sequence of learner attempts, model responses, hints, tool calls, retrieved resources, changes in system state, teacher interventions, or other events occurring during an interaction.

Zhong and Zhu (2026) propose trace-based evaluation in which agent runs become auditable packages containing evidence about execution and verification. An educational analogue could support evaluation of the process through which assistance was provided, rather than limiting evaluation to the quality of individual model responses.

Substantial questions would accompany such an approach. Student interaction traces contain sensitive information. More data collection does not necessarily produce better educational evidence, and increased observability for developers or institutions can reduce privacy and learner autonomy. Decisions about what should be observed, by whom, for how long, and for what purposes would therefore need to be treated as both technical and governance questions.

Nevertheless, observability illustrates why the harness layer may be relevant to educational evaluation. If important properties of an AI-supported learning experience arise from sequences of interactions rather than isolated outputs, evaluation infrastructure may need to operate at the level of the system rather than the model alone.

Implications for AI Infrastructure in Education

AI harnesses are an emerging technical concept, and their significance for education has not yet been established. The terminology may also change as agent architectures mature. Nevertheless, the concept highlights an increasingly important distinction between the capabilities of foundation models and the behavior of systems built around them.

For educational AI, this distinction may be particularly consequential because general-purpose measures of AI performance do not fully represent educational goals. Accuracy, task completion, efficiency, and model capability remain relevant, but learning also depends on processes such as practice, feedback, explanation, reflection, and appropriately calibrated support. Importantly, many the mechanisms governing these processes reside outside the foundation model.

An AI harness can manage context, tools, state, permissions, orchestration, evaluation, and observability. In educational applications, these mechanisms could potentially coordinate models and other infrastructure around specific instructional and pedagogical purposes while providing points of control for learners and educators.

The resulting research agenda extends beyond whether foundation models can produce educationally appropriate responses. It includes questions about how AI capabilities are assembled into learning systems, how pedagogical intentions can be represented without unnecessarily constraining educators or learners, how system behavior can be evaluated across sequences of interaction, and which components are sufficiently generalizable to become shared infrastructure.

The development of increasingly capable models will continue to expand what educational AI systems can do. Harness engineering draws attention to a related problem: determining how those capabilities are organized, constrained, observed, and connected to educational purposes. For AI infrastructure in education, that surrounding layer warrants increasing attention.

References