Part 1: AI Harnesses for Education: What is this Emerging Layer of AI Infrastructure?

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

In an age where budgets are tightening and there is more work on an educator’s plate, there are clear areas of promise where Artificial Intelligence (AI) could automate processes and lighten educator’s loads. However, the inconsistency in the reliability, accuracy and privacy of AI models, as well as the lack of transparency in how they make decisions, are critical barriers to implementing and scaling this technology effectively. AI harnesses are one potential mechanism to transition AI from one-shot prompting and answers to repeatable, reliable, and auditable educational workflows.

What is an AI harness (or an Agent Harness)?

An agent harness is a piece of software that provides an environment for an AI model to operate within. Unlike most AI models, you, as an end user, can own your own agent harness.
– Earendil Product Post What is a Harness?

An AI harness is the surrounding technical architecture and underlying engine that mediates how an AI agent detects an environment, executes actions, collects feedback, and determines task completion (Cabral & Furgale, 2026; Zhong & Zhu, 2026). Practically speaking, completing a task with AI requires coordinating three things:

1. Using an AI model for reasoning (technically, “inference”)
2. Providing context to the model that is task-specific and goes beyond what it already knows, such as user preferences, prompts, reliable external knowledge sources, instructions for how to reason about the specific task, and available tools, functions, and servers that can support reasoning.
3. An overall process loop that controls how the model iterates with the context in order to go from information gathering to producing an output.

Rather than treating AI performance as a property of the model alone, a harness acts as a distinct governing layer that translates foundation model capabilities into controllable, practical system behavior. Some of the harness functions operate at a user level, while others need to be addressed at the system level and both need to be leveraged together to improve performance for education-specific purposes.

In short, an agent engages in an interactive execution framework known as an agentic loop. While the model provides raw reasoning, it relies on context to direct that reasoning. The agentic loop then operationalizes this context, creating a feedback cycle where the model evaluates information, selects an action, processes the execution results, and identifies what to do next to move from information gathering to a final output. Each step in that loop requires infrastructure to manage the context, route tool calls, handle failures, and decide when to stop – that infrastructure is provided by the harness.

Table 1 describes some AI harness functions in the context of a hypothetical educational workflow. For illustration, we imagine an “update my lesson plans” workflow that an educator could trigger at the beginning of every instructional unit they teach, which would examine last year’s lesson plans to suggest a short list of potential improvements. The improvements might come from new district-level documents, a library of lesson plans that the educator finds helpful, or from other sources.

An agent may receive information, select an action, call a tool, process the result, update its state, and determine what to do next. The harness governs which information and tools are available at each step, how state is maintained, and when the process should continue, stop, and require human intervention. These functions are particularly important within an educational context where information quality, data security, and transparency are paramount.

Table 1: Harness Functions

Harness Function	Primary Scope	Education-Specific Examples
Context Management	Information Selection	Organizes lesson plans, curriculum materials, examples of student work, records, and prior interactions input into the model.
Tool Access	External Integration	Dictates which district systems, databases, search engines, or specialized resources the model can access and how those resources are integrated.
State & Memory	Data Persistence	Determines what information, preferences, and context persist across individual actions, interactions, or sessions. Stores the content needed for the present work, (e.g., updating lesson plans for a curriculum unit that begins on a specific date) as well as storing results of previous updates and other information the workflow needs to operate well.
Orchestration	Workflow Logic	Sequences a complex set of steps to do this work, eliminating the unreliability that results from an AI model “re-inventing” a reasoning process on every run, and ensuring that the reasoning follows the teacher’s requests and preferences.
Permissions	Access & Governance	Asks the teacher for permission (and systematically blocks some permissions) before taking any action that may be unsafe, violate privacy, or otherwise dangerous). One example is that the teacher may wish to re-use a prompt from another teacher and the governance layer should scan this prompt before re-use for any dangers and ask the teacher to do the repairs that are needed before the prompt can be used safely.
Verification	Quality & Safety Control	Evaluates, checks, and filters outputs so that the teacher’s request is honored and that the generated output is explained.
Observability	Audit & Monitoring	Maintaining a detailed list of actions taken to enable fixes. This step enables auditing what the system did and did not do correctly. It allows educators to closely monitor and reject actions taken by an AI tool.

Importantly, these mechanisms are not individually new. Software systems have long incorporated permissions, workflows, state, logging, and verification. Educational technologies have similarly incorporated learner modeling, activity sequencing, feedback, scaffolding, and instructional decision-making. Intelligent tutoring research, for example, has long examined how differences in system design shape when and how learners receive feedback (VanLehn, 2011). However, successfully integrating these mechanisms into educational workflows, especially those of districts who are leveraging AI, has been slower than general model adoption

Affordances

An AI harness can provide the functional framework necessary to manage multi-step, agentic workflows across systems and tools. For example, in an education setting, the harness layer controls what instructions, student records, curriculum materials, or prior interactions are fed into the AI model (or agent) and is kept in context and memory across sessions. When it comes to what tools are used, the harness determines which external applications, databases, or search tools the model can invoke, while orchestrating the AI agent calls, retries, and execution paths. For example, a harness layer can ensure that teachers only access their own classroom data, while a principal has access to data across the entire school.

One important advantage of a harness is that it enables system activities, decision steps, and interaction logs to be inspected, evaluated, and audited over time. This harness layer can lead to educators and system designers to map specific instructional, pedagogical, and operational goals directly onto the system that they are using. Allowing them to move beyond generating generic outputs, while allowing for more precise control, gating, and security of the data that enhances the outputs and allows for individualization.

Disaffordance and Guardrails

A harness defines boundaries of what an AI agent cannot do and implements checks to prevent unintended behavior, but reliability and safety cannot be completely guaranteed or fixed. In this way, a harness and using a chatbot are similar; there are guardrails but they can be imperfect. Another limitation is that current model benchmarks evaluate raw capabilities of models in isolation, making them unreliable predictors of real-world safety or system constraints. The AI harness adds a layer that actively governs and dictates how the agent will perform and in doing so, it does add a few guardrails including permission control and verifications.

An AI harness can restrict model access based on user roles and it can also enforce strict rules on which actions each role is permitted to perform. Beyond that, the harness can check model outputs before execution. This allows the harness layer to trigger stopping conditions that can require explicit human authorization for high-stakes actions.

What does this look like in a school district?

Kris Hagel at Peninsula School District describes a scenario in which a principal asks a district agent to prepare for a meeting. Completing that task requires grounding in instructional essentials, the district’s strategic plan, and other initiatives including Universal Design for Learning. Additional information from attendance, grades, assessments, interventions, teacher notes, special education systems, communications, and other district sources may also be required. The model may be capable of summarizing those records, but the surrounding system must determine which records it can access, whose permissions apply, what information persists, and what actions require human review.

Hagel describes applications that can safely participate in district-agent workflows as “agent-ready.” Agent-readiness and harnesses are not synonymous: agent-readiness concerns whether external systems can be safely used by agents, while the harness helps govern how an agent accesses and acts across those systems. Together, they illustrate why implementation depends on more than adding an AI model to an existing product.
The harness concept can be useful because it identifies the surrounding mechanisms as a distinct technical layer in systems built around general-purpose foundation models. Critically, it also shifts the unit of analysis. Benchmarks can tell us whether a model can solve a mathematics problem, interpret speech, analyze documents, or describe images. They cannot, by themselves, tell us what information an educational application will expose to that model, which actions the model may take, or how its behavior will be constrained in practice.

For educational applications, AI models provide a set of technical capabilities, while educators and designers define the instructional purposes those capabilities are intended to serve. The harness provides one layer through which those purposes can be translated into system behavior.

In Part 2, we examine why these harness-level decisions matter for learning, particularly when the objectives of general-purpose AI systems and educational environments do not align. 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.

References