It is half past four. Classes are over, but the teacher's day is not.
One student has made the same mistake on three assignments and needs individual feedback. Another is falling behind and needs different practice. A parent asks about the week's progress. The teaching team wants a summary of the class's common errors.
Tomorrow's lessons still need preparation.
When education discusses AI, lesson plans, question generation, and presentation slides often come first. Those capabilities are useful.
But conversations with classroom teachers reveal that content is not always what they lack.
They lack time, particularly the time needed for continual tracking, repeated organization, and individual feedback.
A teacher may quickly see what one student does not understand. Remembering where forty students have struggled over a month is much harder.
That is one of the most promising directions for agents in education: helping with personalized support that otherwise requires large amounts of teacher time.
A test offers more than a score
Imagine a class of forty students after a mathematics test.
The most accessible results are scores: one student has 82, another 76, another 91. Those numbers alone provide limited guidance for teaching.
More useful questions concern patterns. Which kinds of problems has a student missed on the last three tests? Is the difficulty conceptual or computational? Is the entire class losing marks on one topic?
Teachers can analyze this themselves, but doing it for every assignment and assessment requires substantial time.
In ZGI, this can be designed as a workflow. When assignment or test results arrive, errors are organized according to predefined topic categories. A model can then help prepare class-level and student-level summaries.
The goal is not to label students. It is to organize information the teacher would otherwise have to aggregate manually.
A summary might note an increase in errors on fraction word problems, or flag that a student has struggled with the same concept three times and may benefit from reviewing it in the next practice session.
The teacher decides how to teach in response.
AI can help a teacher see a problem sooner without judging the student on the teacher's behalf.
Personalized learning requires more than generating a question
Language models can readily generate exercises. The harder question is which exercise a particular student needs.
One student may have just grasped linear equations while another is ready for combined application problems. Giving everyone the same generated worksheet is not meaningful personalization.
A better approach uses recent practice results, current understanding, and course goals as context.
A practice agent might query recent assignments, identify difficulties, use a question bank or knowledge base within the teacher's defined scope, and prepare suitable exercises.
New results then inform the next round.
Instead of generating ten questions once, the agent supports a continuing cycle: learning, practice, feedback, and adjustment.
That is the distinction between an agent and a one-off content generator.
Conceptual illustration: organize learning information, let teachers review it, and adjust practice in a feedback loop.
A teaching knowledge base should contain more than textbooks
Textbooks are only part of the knowledge that makes a teaching agent useful.
Teachers' experience matters too: common misconceptions, typical errors, explanations that work at different levels, and when to stop explaining and let students practice.
Much of this is not fully documented in textbooks. It comes from years of teaching.
Schools and education providers can organize curriculum standards, textbooks, teaching research, question banks, answer guides, and internal methods in a shared knowledge base, with access appropriate to each role.
Student questions can stay within the course scope. Teachers preparing lessons can retrieve internal materials and previous examples.
Boundaries are necessary. A student's question should not automatically produce an answer far beyond their stage of learning. Internal teacher resources may not be suitable for direct student access.
Enterprise AI needs to consider who should see knowledge as well as whether it exists.
Parents usually need something more useful than a long AI report
School-home communication can absorb significant time.
A parent asks how a child has been doing this week. The teacher knows, but preparing an individual account for dozens of students each week is demanding.
An agent can organize the information first. Based on assignments, assessments, and classroom records, it can draft a short account of recent progress, topics that need attention, and possible priorities for next week.
We do not believe these messages should be sent entirely automatically.
Educational communication depends on tone, relationships, and an understanding of the student. A teacher should review and revise the draft before it is sent.
Those minutes of human review matter. A system can organize data, but a teacher may understand whether a child has had a difficult week or is facing a longer-term challenge. A few scores do not establish that distinction.
Schools may need less duplication rather than more AI apps
Educational institutions can encounter the same fragmentation as businesses.
The teaching research team connects one model. Marketing uses another tool. Teachers register for several services. Question banks, lesson plans, and student data end up across different platforms.
At first, everyone becomes more productive. Eventually, the institution must manage models, knowledge, permissions, and data together.
This is why ZGI focuses on Agent Runtime rather than one fixed educational application.
Models, knowledge bases, Skills, and workflows can be managed in one environment, allowing an institution to design agents around its needs.
A teaching research agent, lesson preparation agent, internal knowledge assistant, and assignment analysis workflow can reuse the same underlying model and knowledge capabilities.
The result can be a system whose capabilities accumulate rather than four isolated tools.
Why source access and self-hosting matter in education
Student data makes this a necessary question. Names, grades, learning records, and assignments all require careful handling.
Institutions should consider where data is stored, who can access it, which models are called, and whether execution is traceable, alongside the feature list.
ZGI's available source code and self-hosting support can help schools, training providers, and corporate learning teams with suitable technical capacity build an Agent Runtime within their own environment and choose models and data connections according to internal requirements.
Source access is not simply another way to say “free software.” It means the operating system can be understood and adapted.
Self-hosting also requires deployment, operations, and security capabilities. Not every school should maintain its own AI infrastructure.
For basic lesson-plan generation, mature SaaS may be a better fit. Runtime becomes more relevant when AI connects student data, internal knowledge, teaching processes, and continuing operations.
Skills can preserve an institution's teaching methods
A school's core assets include its own teaching methods, not just its textbooks.
How should a difficult concept be explained? How can a teacher distinguish misunderstanding from carelessness? When is more practice helpful? When should practice stop so the concept can be explained differently?
Traditionally, much of this has been shared informally among teachers.
Some explicit and reusable parts may be captured as Skills. An assignment diagnosis Skill can define an analysis sequence. An error attribution Skill can limit acceptable evidence. A feedback Skill can guide language for different ages. A lesson preparation Skill can require alignment with course goals and prior knowledge rather than extending the scope indiscriminately.
Models may change several times a year. Decades of accumulated teaching methods should not disappear whenever the model changes.
That is one of the most promising possibilities for Skills in education.
Be clear about what AI cannot do for teachers
Education is particularly vulnerable to exaggerated AI narratives, including the idea that a strong enough model can become an always-available super-teacher for every student.
Reality is more complicated.
A model does not know why a student suddenly became quiet today. It cannot reliably tell whether repeated mistakes stem from misunderstanding, home circumstances, attention, or emotions.
It should certainly not make definitive judgments about a student's ability or future from a few assignments.
Agents in schools should therefore avoid becoming automatic judges of students. They are better suited to organization, retrieval, drafting, and highlighting signals worth a teacher's attention.
Student evaluation, educational decisions, psychological concerns, and important communication must remain the responsibility of teachers and qualified professionals.
The goal is to give teachers more time to notice and understand their students.
Give teachers more time for the work only they can do
AI can easily produce a lesson plan. A good teacher's value extends far beyond preparing slides.
It includes anticipating where a class will struggle, knowing when to pause, deciding which child needs another question, and recognizing when a score does not tell the full story.
If agents can take on assignment organization, knowledge retrieval, practice generation, data aggregation, and repetitive feedback preparation, teachers can redirect time toward those human responsibilities.
Our hope for educational AI is straightforward: let teachers do less of what machines handle well and more of what only teachers can do.
ZGI's source access, knowledge bases, Skills, workflows, and runtime are infrastructure for pursuing that goal.
People still determine the quality of education.
Top comments (0)