Module 1: AI and Agentic Engineering for Education

Generative AI should be understood as a probabilistic content and reasoning system, not as a database of guaranteed facts. It can transform supplied material, draft explanations, summarize documents, classify text, analyze structured data, create code, and work across multiple modalities when the selected product supports them. Educational value comes from combining those capabilities with authoritative course context and a review process.

Before using AI for teaching work, separate four questions. What is the educational objective? Drafting a worksheet is different from deciding a final grade. What evidence must control the answer? A syllabus, textbook chapter, institutional policy, rubric, dataset, or official standard may be the source of truth. What would happen if the output were wrong? Consequence determines the amount of review. Who has authority? AI can support professional judgment but does not acquire the educator's institutional responsibility.

A useful mental model is capability + evidence + controls + human authority. Capability alone is never enough for a dependable educational workflow.

A chatbot responds to a conversational turn. An assistant adds persistent instructions or knowledge. A workflow deliberately sequences multiple steps. Automation executes a predefined process with limited intervention. An agent can select tools or continue through multiple bounded steps toward an objective. A multi-agent system delegates different responsibilities to specialized agents.

Use the least complex architecture that solves the problem. A one-paragraph classroom explanation usually needs only a well-constructed prompt. A semester-wide course assistant may benefit from approved files and persistent context. A grading or analytics workflow needs explicit evidence, criteria, verification, and approval. Multi-agent orchestration is justified only when decomposition creates real value—for example, one role plans, another drafts, another audits evidence, and the educator resolves disagreement.

Increasing autonomy also increases the need for permissions, state control, evaluation, logging, and stop conditions. Agentic engineering is therefore the design of controlled work, not simply the use of an advanced model.

Modern AI systems may accept combinations of text, documents, spreadsheets, images, audio, video, and connected sources. For educators, this makes it possible to work directly with a syllabus, lecture notes, student-safe datasets, presentation files, diagrams, photographs, or class-performance tables instead of manually retyping the context.

File access does not make a result authoritative. Before using a file-aware workflow, identify which file or source is controlling when two sources conflict; whether the file is current and complete; whether student or employee information should be removed or minimized; whether the model can quote, transform, or calculate from the file without changing its meaning; and which outputs require an independent check.

When images or scanned materials are involved, also verify labels, units, tables, and mathematical notation. A visually plausible extraction can still be incomplete. The same principle applies to spreadsheet analysis: validate the schema before interpreting the statistics.

This course teaches durable principles first, but participants should also recognize the current product landscape. As of September 26, 2026, major platforms such as ChatGPT, Gemini, NotebookLM, and Claude offer overlapping combinations of file analysis, multimodal work, research, customized assistants, and connected tools. Feature availability varies by plan, account, region, and organizational settings.

• ChatGPT is a useful primary demonstration environment for conversational work, projects/files, data analysis, customized assistants, and connected workflows where available.
• Gemini provides multimodal and research-oriented workflows; Gems can preserve specialized instructions and knowledge, while Deep Research can work from selected sources.
• NotebookLM is useful for source-grounded study and synthesis around supplied materials, with citations back to those sources.
• Claude supports document-intensive work and tool/connectivity patterns, including integrations and connector-based workflows.

Because product surfaces change quickly, every platform-specific demonstration should be dated and secondary to the durable workflow principles taught in the module.

AI output can fail while sounding confident. Common failure modes include fabricated citations, unsupported explanations, arithmetic mistakes, stale product information, omitted exceptions, misread tables, hidden assumptions, and automation bias—the tendency to trust a machine-generated result because it appears systematic.

A robust educator should distinguish at least four statuses: SUPPORTED for claims grounded in authoritative evidence, INFERRED for reasonable interpretations not explicitly stated by the source, NEEDS VERIFICATION for material claims that should not yet be used, and INCORRECT OR CONFLICTING when evidence contradicts the output.

Ask the system to expose uncertainty, but do not rely on self-reported confidence as proof. Verification should target the actual risk: check citations against the source, recompute numerical results, compare a grading recommendation with the rubric, or inspect whether a generated lesson really addresses the intended outcome. The stronger the consequence, the stronger the independent check.

Educational AI should be controlled according to the consequence of error. Low-consequence drafting—brainstorming examples, rephrasing instructions, or generating alternative discussion questions—can usually use a light review. Moderate-consequence work such as lesson sequencing, assessment-item drafting, or analytics interpretation needs stronger evidence and checking. High-consequence work such as final grading, disciplinary decisions, certification, or decisions based on sensitive learner information requires direct human authority and institutional safeguards.

A practical approval model is: AI drafts or analyzes; evidence and criteria are made visible; a verification step checks correctness, completeness, fairness, and policy; the educator accepts, modifies, or rejects the proposal; and only the reviewed artifact or decision is released.

Human review is not a ceremonial final click. The reviewer must be capable of changing the outcome and must receive enough evidence to make that decision.

For a full professional-development route, allocate approximately 75–90 minutes to this module: 25–30 minutes for concepts and demonstration, 30–40 minutes for guided practice, and about 20 minutes for the hands-on lab and review. In a condensed workshop, use 45–60 minutes by demonstrating one capability comparison, one hallucination/verification case, and the human-in-the-loop design exercise.

Participants should leave with three artifacts: a personal AI-use boundary, a capability-to-task decision map, and a verified workflow for one real teaching task. These artifacts become inputs to later modules rather than isolated exercises.

• 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.

• Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST. Generative-AI risk profile covering information integrity, privacy, evaluation, and oversight.

• ChatGPT Capabilities Overview — OpenAI Help Center, current capability overview.

• Use Deep Research in Gemini Apps — Google Gemini Help, source-selection and research workflow.

• Get started with Gems in Gemini Apps — Google Gemini Help, customized assistants and knowledge files.

• Claude can now connect to your world — Anthropic, integrations and connected-tool workflows.

• DepEd Order No. 003, s. 2026 — Foundational Guidelines on AI in Basic Education — Department of Education, Philippines, human-centred AI policy for basic education.