Since the wide release of generative AI (genAI) models over the past few years, the notion that AI is affecting our students’ cognition has been pervasive. While this remains a concern, there have been a lot of productive discussions around genAI, and resources to assist with assignment redesign and open conversations with our students are widespread. Enter agentic AI: just as we’ve begun to thoughtfully integrate genAI into how we teach and how our students learn, AI companies have expanded AI abilities, throwing a wrench into our understanding and prompting new questions about what AI can do and how it will interfere with our students’ learning. 

The difference between genAI and agentic AI is based on the scope of action. GenAI is task-focused and requires frequent human intervention via prompting. GenAI output is pattern-based: these models are trained on sets of data and then generate content that looks like that data. In that vein, genAI is constrained, both by what we ask of it and by the chat interface. The AI model might provide you with external links to resources, but it doesn’t have the ability to take any action beyond conversing with you. 

Agentic AI can work beyond the parameters of a chat conversation: it is goal-based, rather than task-based, and it is autonomous. It can work through unpredictable situations, make decisions, and react purposefully, just as we do in everyday contexts. Agentic AI can work independently to complete a series of tasks with very little human input needed. It can take control of your browser, your desktop, or an application, and complete a series of tasks for you. 

What does this look like in practice? Let’s say you lead a faculty research group and need to schedule a series of Zoom meetings. If you’re using generative AI, you could ask ChatGPT for suggestions on polling tools to get input from your group members on their availability. Once your group members complete the poll, you could provide the poll results to ChatGPT and ask it to select a time when all members are available. You could also ask it for help with drafting an introductory email to send to all of the group members. The tasks of creating the Zoom link, creating and sending calendar invitations, and sending the composed email would be up to you.

Agentic AI would be able to handle all of these tasks for you. If you give an AI agent access to your Outlook, it could search for a date and time during which all group members are available. It could create the Zoom link and send the invitation and introductory email all on your behalf. It could even be set up to respond to emails that come in from your group members.

To some, this may sound like an appealing time-saver, but it’s important to think about the reality of the information you’d be providing to the agent. Agentic AI can save time completing tedious tasks, but it requires significant access to do so. In the rise of genAI, we have seen a lack of sufficient policy around privacy protections of the information we share. The stakes around data governance and security with agentic AI are similar, meaning there are huge holes in our ability to control what data AI models store, how that data is used, and the likelihood of data being leaked. OpenAI has acknowledged that prompt injection, a malicious process by which a hacker might target agentic AI to leak sensitive information, is an ongoing problem that may not be solved “for years to come.” Now, imagine that you’ve given an AI agent access to your entire browser–including your password keychain, your inboxes, your D2L, and more. 

But being informed about the risks (as well as potential beneficial uses) for agentic AI is only one component of the ongoing discourse. Perhaps most important to the conversation are concerns around student usage. 

Discourse around student use of agentic AI has been omnipresent, and for good reason. Just as we’ve begun to get some footing around generative AI, agentic AI models have caused a swell that seems big enough to knock us off balance. Questions regarding exactly what an AI agent can do within D2L have been floating around, and the hope that we might be able to do something to “AI-proof” our courses to maintain the integrity behind teaching and learning illustrates the need for clarity around what solutions we have available to us. 

To try to address these questions and concerns, we in the Center for Teaching and Learning decided to test some agentic AI for ourselves. We compiled a variety of D2L assignments, including Submissions, Discussions, and Quizzes and integrated tools like Panopto and Hypothesis. To maintain data security, we used laptops that had been wiped of all data and accessed a version of D2L intended solely for testing purposes (with no student or instructor data stored on it). We also used fake student accounts to access our testing site. In having both Claude Cowork and ChatGPT’s agent work through our test course, we were not surprised to get confirmation that agentic AI can complete all of these types of assignments. 

When agentic AI completes coursework, it navigates through a course site item-by-item and reasons its way through assignments. You can see its “thought” process as it charts the steps it will take, and if it runs into obstacles, it will try different strategies to achieve a goal. As we watched the AI agents work through our test course, we observed familiar characteristics of AI-generated content: discussion posts that felt overly formal and enthusiastic, and reflection responses that were superficial. Overall, the work it completed was passable. While we received some pushback about academic integrity from each agentic AI, we were able to find ways to continue to direct them to complete work. Somewhat deflated, but more equipped with hands-on experiences, we wondered how to turn our findings into actionable information.

Does the presence of agentic AI mean online teaching is gone for good? No, but it does mean we will need to continue conversations with our students about AI ethics and focus on cultivating students’ authentic, personal connections with course content. Asking students to reflect on their learning experiences, recognize metacognition, and synthesize meaningfully are still strategies to encourage original learning and likely the most effective ways to measure achievement when considering whether or not AI was used. 

In terms of detecting AI use, there may be red flags we can see in the Class Progress tool or Quiz Event Log: if a student has moved rapidly through course activities and readings, this might be an indication of agentic AI. However, this kind of detection (similar to Turnitin’s AI checker) should only be used to start a conversation with the student. It is not a fail-safe indication of AI use. Whether or not agentic AI is used, speeding through readings is an engagement issue that is worthwhile to discuss, so framing these conversations around authentic and meaningful engagement with course content is key. Likewise, while safeguards like Start dates on Content Modules, Lockdown Browser on Quizzes, and Turnitin’s AI checker may be helpful stopgaps, they are not infallible methods (e.g., just as there are ways to “humanize” writing to evade Turnitin detection, there are ways to instruct agentic AI to change its pace to evade timestamp detection). 

You may now be wondering, if there is no clear way to tell if students have used agentic AI and no reliable way to stop them, what can we do? The real answer is that there is no singular solution to eliminating the threat of agentic AI, but there are ways that we can advocate for mitigation strategies:

• You can draft open letters to AI companies, asking for their cooperation in programming agentic AI models (see Anna Mills for inspiration). 
• You can share statements written by prominent educational institutions, like the one written by the Modern Language Association.
• You can create and browse AI-aware course redesign efforts. Tim Mousel has an open Google form where you can share and view submissions.
• You can join efforts and conversations hosted by LMS companies. D2L has a number of community groups you can join to engage in discussions around stopgaps and best practices.
• You can participate in ongoing clinics and webinars dedicated to discussing agentic AI, like those offered by the American Association of Colleges and Universities.
• You can join existing collaborative efforts to put pressure on AI by publicizing agentic AI’s willingness to complete course work, as seen in the CHEAT Benchmark Manifesto.

In a post on her Substack, Anna Mills outlines a number of other ways we might be able to use our voices to add to the ongoing dissent around agentic AI. While we may not be able to solve the agentic AI problem in the immediate, and while that is no doubt frustrating, it does not mean that we should lose hope in our students or our online course offerings. Instead, we should put the focus on awareness, open dialogue with students, and thoughtful assignment revisions. 

We may not be able to “AI-proof” our courses entirely, but we can use this moment to connect with students over shared concerns, AI ethics, and the importance of meaningful, authentic learning.