YUZHHNext week's topic is limits. The instructor has the lecture notes, but an old diagram is unclear, exercises need reordering and a short explanatory video needs a script, captions and a formula check. These jobs depend on one another. A request to generate slides captures only part of the work.
An education AI agent, as used here, reads course materials, decomposes a task, calls tools and delivers teaching materials within goals and permissions set by a teacher. A chat window carrying an education-agent label may offer only some of those capabilities.
As of October 23, 2026, public product and policy material points to several developments. Evidence of improved learning needs a separate assessment. The recent examples below relate to teachers' work and lead into a hypothetical lesson-planning exercise.
How AI is moving closer to teachers' daily work
In April 2026, China's Ministry of Education and four other departments issued an AI and Education Action Plan. It calls for intelligent learning companions, teacher support, multimodal resource generation and collaborative lesson planning. Each school still has to assess its own course resources, staff support and acceptance criteria.
In July, Anthropic introduced Claude for Teachers for verified US K–12 teachers. It connects Learning Commons curriculum standards and resources with teaching tools. The announcement describes using Claude Code and Cowork to process material and schedule recurring tasks. Its geography, grade levels and data arrangements differ from those at Chinese universities.
On October 5, Google marked World Teachers' Day by inviting ideas about classroom technology from teachers in collaboration with TED. This was a call for teacher input into design. Alongside the developments above, it suggests growing attention to curriculum, teacher participation and everyday conditions of use. It did not report an agent launch or a learning-outcome study.
Lesson planning, tutoring and administration require different safeguards
Drafting an editable lesson plan, guiding a student through a problem and changing an administrative record carry different consequences if they go wrong. The following are checks for a pilot. Availability depends on the product.
| Application | Suitable tasks | What people should verify |
|---|---|---|
| Teacher preparation | Organize specified material, propose slide structure, draft a script and draw concept diagrams | Accuracy and difficulty of the material, including agreement between illustrations and formulas |
| Student tutoring | Offer course-based hints, follow-up questions and exercise feedback | Whether hints give away the answer and whether the student can work independently afterwards |
| Course resource maintenance | Compare versions of lecture notes and identify affected slides or video segments | Version consistency, citations, permissions and continuing use of outdated material |
| Administrative support | Look up public policies and draft responses to routine questions | Identity, permissions and current information. Changes to enrollment or grades require strict authorization. |
For a school beginning a pilot, teacher planning has definable inputs and deliverables and allows review before class. Student tutoring is also valuable, but satisfaction with an answer and actual understanding require different measures.
Try it on a lesson about limits
This hypothetical design exercise describes neither a customer implementation nor an observed learning outcome.
Specify what students should learn
Supply the approved textbook chapters, notes and exercises. State the goals: students should distinguish approaching a point from attaining it, explain a limit with a graph and recognize when the function value differs from the limit. Include the time available, prerequisites and indispensable derivations.
These constraints give the agent more to work with than a request for high-quality slides. Without them, an attractive outline might misplace the central concept or rely on a proof students have not yet learned.
Review the structure before producing each format
The agent first proposes a sequence, key examples and corresponding source locations. After teacher approval, it prepares editable slides, diagrams and a short video script. A change in teaching order can then be made before narration and video have to be redone.
Every format serves the same learning objective. An open circle may indicate an excluded point while a caption calls it the function value there. Each file could look complete but together they would mislead students. Check consistency across materials beyond simple file delivery.
Retain sources, revisions and the publication decision
Trace each important conclusion to the textbook or another approved source. Mark additional examples separately and flag unsupported material. When notes change, identify affected slides and script passages so teachers can review just those portions. Teachers select the version used in class.
Yuz EDU, currently in private beta for university instructors, organizes source materials, outline approval, courseware generation and review. We are testing whether materials follow the teaching objectives and teachers can check and edit the facts. This does not establish general availability of experimental video features or outcomes validated by a school.
Four engineering requirements
First, limit the sources. Use teacher-approved materials for course answers and generation and preserve their versions and citations. A retrieved passage is only a candidate: check that it supports the claim. Ask for more material when evidence is missing.
Second, save task state. If slides, images, audio or video fail, retry that step while retaining approved work. Network interruptions, unsupported formats and tool errors can otherwise make a successful demo an extra burden in practice.
Third, build review into the process. Confirm the structure, check content and diagrams, then review before release. Earlier checks catch issues that are harder to find across a whole batch. Automation can flag missing formulas or inconsistent captions, while instructors supply subject judgment.
Fourth, narrow access. A course-content task does not require class records. An answer-drafting tool does not need permission to change grades. Handle student work, scores and identity information under school policies and applicable requirements, with authorization, retention limits and access records. Treat model and tool output as data rather than executable instructions.
Count rework when evaluating a pilot
Time to generate a file is easy to measure. Time spent finding and correcting errors often goes uncounted. Compare chapters of similar difficulty within the same course under the current and assisted workflows. Treat small trials as process experiments before drawing broader conclusions.
- Total teacher time. Include preparation, instructions, waiting, checking, editing and publication. Record rework separately.
- Content quality. Record conceptual and formula errors, incorrect citations and unsupported conclusions. Report serious errors separately from averages.
- Usability. Can teachers edit the slides and diagrams or correct captions? Can they locate affected material after course updates?
- Learning. To evaluate outcomes, observe independent answers and explanations without AI and use a delayed test where possible. Document differences in prior attainment, teaching and question difficulty.
- Operating costs. Count model, image and video calls, failed attempts, retries and additional operations teachers have to learn.
If production gets faster but checking becomes harder, recalculate the benefit. Neater assignments without better independent explanations do not establish a learning gain. These measures say more about the school's problem than a slide count.
Make one lesson reliable first
Teachers need materials they can use tomorrow, errors they can fix in time, resources they can reuse next semester and time to support students. AI offers scope to reorganize this familiar work.
For teams including YUZHH, a narrow but complete workflow is a useful first step. Teachers specify sources, the system delivers inspectable and editable material, issues can be traced, and revisions persist. Once that works for one lesson, there is a better basis for extending it to a course.