Author: Pati Ruiz, Jordan Mroziak, Keun-woo Lee, Megan Pattenhouse
In Part 1 of this series, we introduced AI harnesses as the technical layer that helps translate model capabilities into system behavior. For educational applications, this distinction matters because systems optimized for general-purpose assistance may not be optimized for learning. This post examines why harness-level mechanisms may be important when educational goals depend on when, how, and under what conditions AI capabilities are made available.
Research in the learning sciences has long shown that instructional support is not only a question of whether assistance is available, but how it is structured. Formative assessment relies on evidence of learner understanding to inform subsequent instructional decisions (Black & Wiliam, 2009). Intelligent tutoring research has similarly examined when feedback and support are introduced during problem solving (VanLehn, 2011), while work on productive failure has examined conditions under which difficulty before instruction can contribute to later understanding.
This points to a key tension in the design and purpose of AI systems; general-purpose models are designed for task completion, rather than for the process of learningA system may increase the likelihood that a learner produces a correct answer without increasing the learner’s capacity to produce that answer independently later. In fact, unrestricted assistance from GenAI can reduce some of the cognitive activity an instructional task was intended to elicit. Wang and Shan (2026), for example, argue that educational AI architectures should account for the processes that precede a correct answer rather than optimize only for answer completion.
Why the Surrounding System Matters
Consider an AI-supported mathematics activity. The underlying model may be capable of generating a complete solution, a hint, an explanation, or an alternative example. The educational question is not only whether those capabilities exist, but when and under what conditions each should become available.
Rotter, Benazet i Montobbio, and Hernández-Leo (2026) examined this issue by treating access timing as a form of scaffolding. Participants who received strategically timed access to generative AI demonstrated stronger post-test performance and metacognitive accuracy than those with unrestricted access. The study is an early example and should not be generalized directly to K–12 settings, but its architectural relevance is important: the pedagogically consequential intervention occurred in the system governing access to the model rather than through modification of the model itself.
This illustrates a broader implementation challenge. “Preserve productive struggle” is an educational objective, but it is not yet a technical specification. Implementing that objective may require decisions about prior learner attempts, available hints, tool access, timing, escalation, and educator intervention.
Beyond Prompt-Level Instructions
Translating an educational objective into system behavior may require more than prompt-level instructions. For example, a prompt might instruct a model to generate hints rather than complete answers in support of productive struggle. A harness could additionally apply rules based on prior learner attempts, sequence available hints, restrict particular tools, retrieve evidence from prior work, or establish when more direct assistance becomes available.
The same implementation challenge applies to formative assessment. A goal such as “use evidence of student thinking” does not by itself specify which evidence should be available, how it should be represented, which learning goals it should be associated with, what criteria should guide interpretation, or which outputs should be presented to students or educators.
Similar considerations apply to metacognition, learner agency, and teacher orchestration. Harness-level mechanisms could structure opportunities for prediction or self-explanation and use information such as prior attempts, progress through an activity, or previous forms of support to determine which options are available next. Just as importantly, they can preserve opportunities for educator judgment and intervention. Research has consistently associated positive teacher-student relationships with student engagement and achievement (Roorda et al., 2017). Harness design should therefore consider not only what can be automated, but how educators retain visibility into learning activity, receive relevant evidence, and configure, pause, or override automated actions. Without these mechanisms, AI-mediated support could reduce opportunities for educators to interpret learner needs and shape subsequent instruction.
Preserving these key learning processes is a long-standing concern in adaptive learning and intelligent tutoring systems. Harnesses offer one technical layer through which general model capabilities can be coordinated with application-specific instructional mechanisms such as retrieval, state management, tool permissions, evaluation, and orchestration. Doing so requires educational intentions to be represented in sufficiently precise technical forms to guide technological system behavior.
Identifying desirable educational behaviors, however, is only part of the infrastructure challenge. Pedagogical goals must ultimately be represented in forms that developers can implement, evaluate, and integrate into educational products. Part 3 examines that implementation gap and what high-quality educational infrastructure must provide so that downstream developers can put models, datasets, harnesses, frameworks, and other public goods to practical use.