Module 10: Integrated Agentic Teaching System

The capstone integrates the course around one real or representative course, unit, or professional-training problem. It should not become a collection of unrelated prompts. Every artifact must connect to the same intended outcomes and source-of-truth set.

The participant begins with a short system charter covering teaching context, learner group, problem, intended outcomes, governing sources, constraints, success criteria, and human authority.

The charter bounds the system and makes later decisions auditable. If the participant changes the outcomes or source materials midway, the change should be recorded rather than silently propagated.

List every authoritative input: curriculum, syllabus, institutional policy, adopted references, templates, datasets, rubrics, and approved tools. Identify which source controls when two inputs conflict.

The capstone should preserve traceability from source → AI action → artifact → verification → approval. A final lesson plan or grade summary should not be detached from the evidence that produced it.

Use revision identity where material. If the assessment uses Rubric v2, the grading workflow should not accidentally use Rubric v1. If a course policy changes, the affected downstream artifacts should be identified for review.

The first pipeline transforms governing requirements into measurable outcomes, an alignment map, lesson plan, presentation storyboard, learner materials, and activities. It should reuse shared context rather than independently regenerating each artifact from memory.

At each stage, define an acceptance gate. Outcomes must trace to the curriculum; the alignment matrix should have no orphan outcomes; presentation materials should match the verified lesson; accessibility checks should pass; and generated visual claims should be verified.

Only accepted artifacts become inputs to later stages.

The assessment pipeline should start from outcomes and a TOS, not from a request for random questions. It then produces items/tasks, answer keys or expected solutions, rubrics, and validation evidence.

The grading workflow uses the verified criteria and learner evidence to produce a proposed score and feedback. The educator confirms or changes consequential results.

Traceability should allow a reviewer to answer: Which outcome does this item measure? Which rubric was used? What submission evidence supported this score? Who approved the final result?

After assessment, the system may analyze de-identified performance data, item behavior, and outcome attainment using the institution's actual rules. It should separate descriptive findings from causal explanations.

A credible learning gap leads to a targeted intervention and a plan for follow-up evidence. CQI is documented only after the action and subsequent evidence exist.

The capstone should therefore close the loop rather than end at grading: Plan → Teach → Assess → Review → Analyze → Intervene → Reassess → Improve.

The participant may implement the system with ordinary prompts, projects/files, assistants, connectors, or agents. The architecture should be justified by the task.

Document each automated or agentic element using its objective, inputs, tools, permissions, output, stop conditions, verification, and human approval.

Use least privilege. A planning assistant does not need access to official grades. A grading assistant does not need permission to email students. A research assistant should not publish citations it cannot verify. The goal is a defensible workflow, not maximum automation.

A polished demonstration is not enough. Test the capstone with normal, incomplete, conflicting, and adversarial cases.

Required examples include a missing authoritative file, conflicting instruction and policy, unsupported citation, unexpected but valid student solution, malformed analytics data, unauthorized tool action, and tool or connector failure.

For each failure, record whether the workflow stops safely, requests review, or produces an incorrect artifact. Correct the responsible layer and rerun earlier tests. This creates a small regression suite for the teaching system.

The final evidence package should contain: a system charter and source-of-truth map; T-E-A-C-H prompt/workflow specifications; instructional plan and alignment map; presentation/materials workflow; assessment blueprint/TOS and sample assessment; grading and feedback workflow; analytics/intervention/CQI workflow; agents/tools/permissions map; privacy, integrity, and governance controls; and evaluation cases, failure log, limitations, and maintenance plan.

A recommended 100-point capstone rubric is: Instructional alignment — 15; Teaching-material quality — 10; Assessment validity — 15; Grading and feedback design — 15; Agentic workflow design — 15; Verification and reliability — 15; Privacy, ethics, and governance — 10; Documentation and traceability — 5.

The rubric is a course design recommendation until CE implements a formal course-level competency-certificate workflow.

The complete course is suitable for approximately 12–15 hours of combined instruction, guided practice, module labs, and capstone work. A condensed instructor-led route can fit approximately 8–10 hours by using one flagship demonstration and one compact lab per module, with the capstone presented as a guided integration exercise.

CE's current assessment system can issue completion certificates for eligible completed 20-question quizzes. A future course-level competency certificate should be treated separately and could require all modules, topic assessments, the practical activities, and a reviewed capstone. Until that feature exists, the lesson must describe it as planned rather than already active.

• Guidance for generative AI in education and research — UNESCO. Human-centred guidance for generative AI in education and research.
• Artificial Intelligence Risk Management Framework (AI RMF 1.0) — NIST. Framework for governing, mapping, measuring, and managing AI risk.
• Model optimization — OpenAI. Evaluation-first guidance for testing and improving AI outputs.