1. The Real Problem

AI misuse in education no longer stops at copy-pasting answers from chatbots. It has taken one step further: agents can operate autonomously instead of just answering when asked.

Blackboard — one of the largest LMS platforms — has warned that AI Agents are capable of taking over nearly the entire student learning process, from reading materials and completing assignments to submitting results, making academic integrity assessment nearly impossible without new monitoring mechanisms.

A specific example of the “browser that acts on behalf of users” trend is Comet — Perplexity’s AI browser, launched mid-2025 and made freely available on all platforms from early 2026. It’s important to clarify: Comet is not an “AI model” but a browser with integrated agent capabilities, able to open tabs, read page content, fill forms, and complete multi-step tasks on behalf of users. It wasn’t designed specifically for “taking exams for students,” but this agentic-browsing capability is exactly the type of tool that concerns educational platforms: the agent doesn’t need to know the exam structure in advance, just how to read and click like a real person.

The most obvious consequence: grades can increase without reflecting actual ability. The problem isn’t “lazy students” but that tools have outpaced what evaluation processes can keep up with.

2. Practical Experiment

To understand how agents work at a basic level, I set up a simple environment: Ubuntu 22.04, Python 3.10, no GUI — just terminal.

An important note to clarify: the openai library has completely changed its syntax since v1.0 (released November 2023). The old way of calling it (openai.ChatCompletion.create(...), openai.api_key = ...) no longer works with current SDK versions — it will throw errors immediately. The correct syntax now uses a client object:

$ pip install openai
$ export OPENAI_API_KEY="sk-..."
$ python - << 'EOF'
from openai import OpenAI

client = OpenAI()  # automatically reads OPENAI_API_KEY from environment variable

response = client.chat.completions.create(
    model="gpt-5-mini",
    messages=[
        {"role": "system", "content": "You are an educational assistant."},
        {"role": "user", "content": "Explain the Pythagorean theorem in simple terms."}
    ]
)
print(response.choices[0].message.content)
EOF

I need to be direct: the code above is not yet an agent. It’s just a single chat API call — ask a question, get an answer, done. A real agent needs at least two additional things: (1) the ability to call tools (tool use) — like reading files, accessing the web, manipulating the browser; and (2) a loop that allows it to decide the next step based on the previous step’s result, instead of stopping after one answer. If you want to build a real agent, you can start with frameworks like LangChain or CrewAI, which provide ready-made tool-calling mechanisms and decision loops.

The key takeaway from this small experiment: the gap between “chatbot API call” and “agent acting on behalf of users” is not small in technical terms — but for end users, that boundary is nearly invisible. That’s exactly why educational platforms find it difficult to detect.

4. Solutions

A ban is not a sustainable solution — these tools will only become more accessible. Here are some more practical approaches:

  • Design assignments that are hard to “agentize”: require live presentations, explain the thinking process, or use personal/realtime data that agents don’t have ready context to handle well.
  • Let students learn how to build agents, not just how to use them. Understanding that agents need tools and loops to function is also understanding their limitations — and those limitations are exactly where teachers can create exam questions.
  • Verify the process, not just the result: log operations, short explanatory videos, or assignments with mandatory intermediate steps to submit.

AI Agents in education are neither automatically a threat nor a benefit — it depends on whether evaluation processes are redesigned to account for their existence.

Sources: OpenAI official documentation on Python SDK v1; Blackboard/Anthology warnings about AI Agents in LMS; Perplexity Comet updates (2026).