Interactive learning material · M1-T2-MLC2

Agents and Automation Workflows

From answers to actions – with the teacher in control

Author: Zsolt LengyeliTStudy Hungary Training and Research Centre
Reviewer: Mária HartyányiiTStudy Hungary Training and Research Centre

The teacher sets a goal for the agent, which delegates tasks to subagents working with files, documents and a browser. — Goal setting → execution → teacher review

Learning objective

A large part of a teacher’s week is spent on multi-step digital routines: collecting and sorting materials, writing similar messages, adapting worksheets, checking who has submitted what. The aim of this learning content is to introduce AI agents and automation workflows in an educationally meaningful way. After studying this material, teachers will be able to explain how AI agents differ from chatbots and identify tasks involving information organisation, routine communication and teaching resource preparation that can be safely supported or partially automated by AI agents.

Introduction · From answers to actions

Introduction: from answers to actions

The chatbot answers

A chatbot answers, explains and makes suggestions. Carrying out the task – such as converting a file or saving a post – often still falls to the user.

The agent also takes action

An AI agent can carry out steps using the tools and permissions available to it: opening files, organising folders, filling in browser forms, running small programs or preparing message drafts.

You set the direction

From now on, it is worth thinking like this: the person using the AI is the boss. An agent – however capable – does not know the goals and priorities of the school, the class or the individual learner. Like a team with talented staff but no clear leader, it works well only if someone knows what to do, delegates clearly, checks the results and corrects course. The quality of the outcome depends on the quality of the boss’s thinking.

Basics · Agents and subagents

What is an AI agent?

An AI agent is an AI system that not only answers a question, but also independently pursues a goal: it makes decisions, plans steps and executes them, using the resources of the computer – files, a web browser, programs and connected services. The decisive difference is the last element: the agent does more than produce text; it actually works on the machine.

Larger tasks are often split up. A main agent acts like a project manager and delegates narrow sub-tasks – for example searching for a document or checking a file – to subagents, task specialists that complete their part and report back. This keeps the main agent focused on the overall goal.

The agent breaks the teacher’s goal into sub-tasks, and the subagents’ results return to the teacher for review. — The teacher sets the goal; the agent plans and delegates sub-tasks to subagents that work with files, documents and the browser; the results come back to the teacher for checking

Basics · Tool examples

Where we meet agents today

Agent capabilities in chat assistants

An agent assistant can use a browser, open websites, collect information and prepare drafts. For example, it can prepare a school blog post and – where the system supports it and access has been granted – save it as a draft.

ChatGPT Work supports multi-step tasks similar to the ChatGPT agent-mode example in the source material; availability depends on the account and the school’s settings.

Browser extensions

With permission, Claude and ChatGPT browser extensions can interpret an open page, summarise its content and perform actions on it. Available features and access permissions vary between tools.

Desktop agent applications

Applications that work with local files and programs – such as Claude Desktop’s agent capabilities or Codex – can manage files and write and run programs using the tools available to them. Browser and screen control may require separate capabilities and permissions.

M1-T2-MLC1 introduces Codex and its permission levels in more detail.

For these tool examples: ChatGPT browser capabilities (opens in a new tab), ChatGPT browser extension (opens in a new tab), Claude in Chrome (opens in a new tab). Checked: 26 October 2026.

Applications · Two types of automation

Rule-based automation and AI agents

A fixed recipe

Automation is older than generative AI. Rule-based workflows – mail rules, scheduled tasks, or no-code tools such as Microsoft Power Automate, Zapier or Make – follow a fixed recipe: “when a file arrives in this folder, copy it there and send me a notification”. They are predictable and repeatable, but not flexible.

Judgement and flexibility

Agents are flexible: they can interpret a messy folder or an unusual request – but for the same reason they are less predictable. Today the two methods are increasingly combined: many assistants can run a prompt on a schedule, and automation tools can call an AI model at a certain step of the workflow.

Applications · Information and communication

Where automation can help teachers

Information organisation

An agent can collect resources – manufacturers’ manuals, safety data sheets, videos, earlier worksheets – into a clear folder structure by topic, rename files consistently and create an index table with a short description of each item. For an electrical training course, for example, a semester’s materials can be sorted into “Circuits”, “Measurement”, “Safety regulations” and “Installation practice”.

Routine communication

Many messages follow the same pattern: reminders to apprentices about required protective equipment, information for parents about a workshop visit, or a note to company mentors about a changed assessment date. An agent connected to the mailbox can prepare these as drafts from a template. The rule is simple: prepare drafts, but do not send them. The teacher reads, personalises and sends each message – a message to a parent or a mentor is part of a professional relationship.

Four teaching tasks: organising resources, drafting messages, adapting materials and preparing scheduled summaries. — Four areas of teachers’ work where agents can help: organising information, preparing routine messages, adapting resources and supporting recurring workflows

Applications · Preparation and summaries

Teaching resources and recurring workflows

Resource preparation

Agents can produce several versions of a worksheet at different reading levels, convert a Word document into a Moodle quiz or a printable checklist, or create parallel versions of a test for a cookery or hairdressing class. Converting and formatting are typical agent strengths – but the subject-specific content (recipes, temperatures, chemical safety, technical values) must always be checked by the teacher.

Workflow support

Recurring tasks can be supported on a schedule: every Monday morning, a summary of last week’s submissions (who submitted, what is missing, common problems), a list of upcoming deadlines for apprentices’ portfolios, or a structured summary of a department meeting, including the agreed tasks. Such summaries save time, but they are a starting point for the teacher’s judgement, not a substitute for it.

Worked example · Weekly submission summary

Worked example: a weekly submission summary with an agent

This example shows how to separate exploration, approval and execution.

Example prompt · Work only on copies

“Work only in the folder Logbooks_week42_COPY. The files are workshop logbook entries from second-year car mechanic apprentices; names have been replaced by codes (Learner A, Learner B …).

Step 1

Explore only: list the files and tell me which learner codes have submissions, which do not, and which files you cannot open. Do not change, move or delete anything.

Step 2

Only after my approval, create one new file named summary_week42.docx with a table (code, task, submitted yes/no, completeness note) and the three most common problems you see, each with one short quotation as an example.

Do not grade anything and do not contact anyone. If something is unclear, stop and ask me.”

Control · Approval checkpoint

Keeping the human in charge

Errors with agents almost always come from two sources: a vague, carelessly written prompt, or a task that we do not understand clearly ourselves. Since the agent really carries out actions, a misunderstanding can lead to modified or overwritten files, deleted work or wasted computer resources (endless loops, unnecessary installations, many paid requests).

The agent can prepare drafts and work on a schedule. Sending, publishing content and making payments require a human review checkpoint.

The agent prepares a document and message; sending, publishing and making payments require human approval. — The agent can work on a schedule and prepare drafts, but sending, publishing content and making payments require human approval

Control · Practical rules

Six practices for keeping control

The following practices keep the teacher in control:

Start with narrow permissions

Agents typically offer three levels: requiring approval before every sensitive action, allowing work with changes submitted for review, or granting full access. Start narrow, observe how the agent works, and only then loosen the reins (least privilege).

Are you asking for exploration or action?

“Only analyse; do not change anything” must be written explicitly.

Work on copies in a separate folder – never on the only copy of important material.

Ask for a draft first

Messages are saved as drafts, posts are saved as unpublished drafts, orders are not placed.

Build in an approval checkpoint

Approval checkpoints are mandatory for sensitive data, payments, publishing, ordering and anything that reaches learners, parents or companies.

Watch and intervene!

Follow the visible work steps and results. If the agent heads in the wrong direction, stop it and clarify the instruction.

Use trusted extensions and apps

A browser extension can see what is on the page, so think about what is open while it runs.

Risks · Human relationships and professional skills

Where does over-automation begin?

Automation is useful up to a point. Beyond that point, the risks grow faster than the time saved:

“You can outsource your thinking, but not your understanding.”

Errors at scale

a mistake by a person affects one message; a mistake in an automated workflow affects thirty messages or every file in a folder – and may go unnoticed.

Deskilling

if the agent always prepares the materials, teachers may lose the habit of thinking them through, and learners may miss out on the practice they need for their trade. “You can outsource your thinking, but not your understanding.”

Weakened relationships

generic, automated messages and feedback are noticed by learners, parents and mentors. Personal contact is not an administrative task, it is a part of teaching.

Responsibility cannot be delegated

the teacher remains accountable for every message sent, every grade and every teaching resource used.

Manipulated content

web pages or documents may contain hidden instructions that try to redirect an agent (prompt injection) – another reason for limited permissions and human approval.

Risks · Data and decisions

Data protection

Use an approved environment for personal data

Do not enter learners’ names, grades, health information or behaviour notes into public consumer AI tools. Use tools approved by the school, with an appropriate data processing agreement, and follow its data protection rules.

Work with fictional data or codes in the examples (Learner A, Learner B). Coding alone does not necessarily anonymise data: if information can be linked back to a learner, it may still be personal data.

Do not let an AI browser extension work while the electronic register (such as KRÉTA) or another system containing learner data is open.

Decisions remain a human responsibility

As a general rule, Article 22 of the GDPR protects people against decisions based solely on automated processing that have legal or similarly significant effects; exceptions are subject to conditions and safeguards.

Annex III of the EU AI Act lists AI systems intended to decide on admission or evaluate learning outcomes among high-risk uses. The specific classification depends on the conditions in the regulation.

Article 4 of the AI Act also requires relevant providers and deployers to take measures supporting the development of AI literacy.

Official information: European Commission – automated decisions (opens in a new tab); AI Act – Annex III (opens in a new tab); AI literacy – Article 4 (opens in a new tab).

Risks · Consider your own task

Should I automate this? Five important questions

Think of a specific, recurring teaching task and answer the five questions. Your answers provide a starting point for deciding how to proceed.

How should you interpret your answers?

If the first two answers are “yes” and the last three are “no”, the task is a good candidate. Otherwise, use AI only for drafts and make the final decision yourself.

Takeaways · Thoughtful task delegation

Takeaways for teachers

AI agents open up new possibilities for teachers by moving beyond answering questions to carrying out practical tasks. They can reduce administrative workloads, streamline routine processes and free up time for teaching and supporting learners. However, using agents effectively requires more than writing good prompts: teachers need to understand what tasks to delegate, how to set clear boundaries and when to retain direct control.

Define the goal and the expected output!
The teacher knows the learners and the professional setting. The agent needs a precise task and clear boundaries.
Start with copies and narrow permissions!
Ask for exploration first, then approve the necessary steps after reviewing the plan.
Follow preparation with human review!
The teacher checks the professional accuracy of drafts, summaries and adapted teaching resources.
Preserve relationships and opportunities to practise!
Personal communication, pedagogical judgement and skills development must remain part of teaching.

Self-check quiz

Check your knowledge!

Choose one answer for each question, then check your results. You will receive a short explanation for each question.

1What is the key difference between an AI agent and a chatbot?An agent always gives an error-free answer.An agent can also perform actions using the tools available to it.An agent only writes longer texts.

2Which task is the best first experiment with an agent?Assigning learners their final grades.Automatically sending emails to parents without review.Planning how to organise copies of files containing no personal data.

3When should you choose fixed, rule-based automation?When the task is fully predictable and can be described by a precise recipe.When every situation requires open-ended judgement.When nobody can check what a correct result looks like.

4What does the weekly-summary prompt require before a new file is created?Deleting the original files in the folder.Exploration followed by the teacher’s approval.Contacting learners directly.

5How should an agent support routine communication?Every generated message can be sent to all recipients automatically.The professional relationship can be handed over entirely to the agent.It should prepare a draft that the teacher reads, personalises and sends.

6What remains the teacher’s responsibility in an automated workflow?Professional review, compliance with data-handling rules and the final decision.Only granting as much access as possible.Accepting the agent’s output without checking it.

Sources · Related materials

References and further reading

References

Lengyel, Zs. (2026): Introduction to the concepts of artificial intelligence; Concrete applications for content generation. VITA training presentations.
UNESCO (2023): Guidance for generative AI in education and research. Paris: UNESCO.
European Commission (2022): Ethical guidelines on the use of artificial intelligence (AI) and data in teaching and learning for educators. Luxembourg: Publications Office of the European Union.
Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 4 and Annex III, point 3; Regulation (EU) 2016/679 (General Data Protection Regulation), Article 22.

Links

Images without a source attribution were generated using Claude for Mac.

VITA · Contributors and attachments

Credits and copyright

Author: Zsolt LengyeliTStudy Hungary Training and Research Centre
Reviewer: Mária HartyányiiTStudy Hungary Training and Research Centre
Project ID: 2025-1-LT01-KA220-VET-000362456

Attachments

Download the source materialM1-T2-MLC2-TEXT-EN.pdf Download the practical activitiesM1-T2-MLC2-EXC-EN.pdf

This work may be reused and adapted under the terms of the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

Funded by the European Union. Views and opinions expressed are those of the author(s) only and do not necessarily reflect those of the European Union or the National Agency.