What Writing Assignments Taught Me About AI Agents

Published 2026-04-17
tags: #agenticAI

When I was teaching my summer 2025 AI course (AI as a Research Partner), halfway through, OpenAI demoed their first AI agents. Of course, I immediately pivoted to include an "AI agent" week in my course, but back then, I really didn't get it. The demo, while impressive, showed an AI agent navigating a web browser, but I just couldn't see why this was helpful or useful. Fast forward today to April 2026, and everyone is buzzing about Anthropic's AI agents (using Claude code), so I decided to give it another look. And it finally clicked. Suddenly, any mindless repetitive task that was getting in the way of doing exciting, engaging, thoughtful work could be automated away. Now, every time I find myself doing busywork, I automatically start brainstorming new ways for AI agents to do it for me. The trick is, to get the most out of them, you need to provide ample context. You need to think of ways that the agent will misinterpret what you want. You have to spell out everything the agent should and should not do. Because if you don't, the agent might make its own decisions and not do the work that you actually want. If you are an educator, maybe this sounds familiar because this is exactly the exercise my brain goes through when making writing assignments for my students. As it turns out, years of making assignment prompts had been quietly training me for the agentic AI era.

Before we get too far, let's start with what AI agents enable that wasn't possible before. In a podcast interviewwith Jack Clark, Anthropic's co-founder and head of policy, Ezra Klein described the Agentic AI revolution like this:

The AI applications of 2023 and 2024 were talkers. Some were very sophisticated conversationalists, but their impact was limited. The AI applications of 2026 and 2027 will be doers.

And what can they do? We've already seen them try to navigate websites. And while AI agents have many features, I want to focus on how they can read and write files directly on your computer if you provide access. These could be plain text files, but also word documents, powerpoint presentations, interactive websites, and more. To give you an example of how powerful this is, I recently used Claude Code to batch edit 8 slideshow presentations for my anatomy course and make them more accessible by automatically fixing reading order errors, labeling titles on my slides, changing font size and color of select elements, and drafting alt-text for all 142 embedded images. What would have taken me hours of mind-numbing clicking, copying, and pasting, instead took about 10 minutes of set up and prompting. After I prepped and sent the agents on their way, they disappeared and completed this work in the background while my brain was free to work on other more exciting projects. The key to getting results like that, I've found, comes down to the same thing that makes or breaks a good assignment: the instructions you write before you hand it off.

The problem with agents is that right now, they're really slow. When they're navigating through your files or different websites, they'll sometimes just wait for a few seconds before executing a task. There are even times when the agents start thinking in the background, making elaborate plans for painstaking minutes debating with themselves on what to do next and how to proceed. In some ways, that's a good thing because the slowness is a byproduct of the agents trying to reason and problem solve. But, because of this limitation, they really work best when they are sent away to independently complete a task until they're done. In that same interview with Ezra Klein, Jack Clark explains prompting agents less like back-and-forth conversation and more like sending a message in a bottle.

So often it's not just knowing what the task is, because you and I could talk about a task to do and you have intuition, you ask me probing questions, all of this stuff. It's making sure that you've set it up so it's like a message in a bottle that you can chuck into the thing and it'll go away and do a lot of work. So that message better be extremely detailed and really capture what you're trying to do.

Listening to this interview, I can't help but see the comparisons between a good task prompt for an agent and a good assignment prompt for a student. In both cases, context and completeness are crucial. In both cases, we're sending a set of instructions and expecting a certain product to be returned after some time. And in both cases, the overall quality of the product is dependent on the transparency you have with your expectations. Think back to the last time a student surprised you with just how much they misinterpreted what you meant. And most of the time, it's not the student's fault. Yes, you should have said "compare and contrast" instead of just "compare" in that essay question. Yes, you should have specified how long you expected their response to be. Educators have learned through trial, error, and experience that clarity in expectations is necessary; that's why we have rubrics. From a practical standpoint, I often think of rubrics as tools to ensure fair and objective assessment, but rubrics also communicate the parameters of the assignment and our expectations to the students as well.

In education, we communicate these details to focus our students' attention and their efforts. We know what we're looking for, and to get the best out of our students, so should they. No one wants to receive or even assign a bad grade based on misunderstanding. And while the stakes might seem lower for providing an AI agent with vague incomplete instructions, it's worth being economical with agentic AI. Because agents spend a lot of compute to reason and problem solve, they use more resources than interactively chatting with the chatbot. AI is starting to become a huge drain on our resources and it's important that we're mindful with our usage. Many users are reaching their daily limits of compute with these models, and for most of us, it's because we're not spending enough time thinking about what we're asking for. If you have to send your agent out 3 or 4 times to complete your tasks correctly because you didn't specify some important parameters in your prompt, that's a real bottleneck. It's the same problem as a student coming back to ask clarifying questions three times before they can even start, which is a signal the assignment needed better instructions from the beginning.

If you want to explore agentic AI and what it can do for you, I think it's important to start practicing with it now. The only way that we learned to improve our assignment instructions was to try and receive feedback or to learn from others' experience. Start small. Pick one piece of tedious, repetitive work and describe it to an agent as thoroughly as you'd describe it to a student. Be specific about the outcome you want, anticipate where it might go wrong, and build in guardrails the way you'd build a rubric. You'll likely have to refine your prompt a few times, and that's okay. That's the same iteration loop we already know from teaching. The good news is that educators have been writing detailed instructions for years. The agentic AI era may just be the moment that skill finally pays off in an entirely new way.