# revision history may not catch the agent Source: https://aaronmakelky.com/blog/agentic-ai-classrooms-what-teachers-need-to-know Last updated: 2026-02-20 Agentic browsers can move through web assignments in ways revision history may not catch. That changes what a teacher can verify after the work is done. [Back to blog](/blog) Field notes February 20, 2026 · 12 min read Classroom AI · Agentic AI I hit "run" on an agentic browser, handed it a Google Doc full of questions about the League of Nations, and walked down the hall to get a drink of water. When I looked through the classroom window, my high school students were staring at the screen while tabs opened and answers typed themselves into the document. A few of them thought I was remote-controlling the computer. I just waved. Watching that, I stopped thinking of AI as a copy-paste shortcut. It could work on its own now. I sat down with Lindy Hockenbary on her [Make EdTech 100](https://makeedtech100.transistor.fm/episodes/education-in-the-ai-agent-era-with-aaron-makelky) podcast right after TCA conference in February 2026 to talk through the latest round of AI releases. We came in with a list of updates and left rethinking what these tools actually mean for classrooms. Prefer to listen? [Catch the full episode with Lindy here.](https://makeedtech100.transistor.fm/episodes/education-in-the-ai-agent-era-with-aaron-makelky) ## A packed room of educators had never heard the word "agentic" At TCA, Lindy was sitting in a session when the presenter mentioned agentic browsers. The room, packed and standing room only, went sideways. Most people had never heard the word "agentic." Some who had heard it immediately started asking how to block it. Agentic AI has been the main thread behind most of the big releases over the past year, yet the majority of educators at a major tech conference were encountering the concept for the first time. Here's the simplest way I can explain it: agentic AI has tools. If you gave ChatGPT a toolkit, a wrench for this job and a hammer for that one, that's what makes it an agent. Maybe it gets a browser and can run a search. Maybe it can write to your Google Doc. Maybe it fires on a schedule. Monday mornings at 8 AM, run this task. Maybe it's a trigger: when I get an email from this person, do this. The brain is the same AI you've been using. The tools make it an agent. Generative AI waits for you to prompt it, copy the output, and paste it somewhere. Agentic AI goes and does the thing. You tell it to respond to your emails saying you're out next week at NCC in Seattle, and it opens your inbox, drafts the replies, and hits send. The autonomy is what freaked people out in that session. ## Revision history is not a safety net anymore If you're an educator relying on anything in a web browser to assess student learning or track progress, you need to hear this: a browser agent can do the assignment for the student, and the revision history will look organic. One way teachers have been catching AI-generated work is checking the revision history of a Google Doc. If a student pasted in a four-page essay all at once, that was the tell. That worked for about a year. It doesn't work now. A browser agent doesn't paste one giant block from a ChatGPT tab. It navigates into Canvas or Schoology or Google Classroom. It shows the student's account logged in. It types at a realistic speed. I've had students show me that you can prompt it to add a couple of typos, then go back and correct them. The whole thing runs without the student touching the keyboard, but the revision history looks like a person sat there and wrote it. There are even Chrome extensions built for this. You paste your AI-generated essay in, set the typing speed, and it goes into your Google Doc character by character. Those extensions are free. They're on the internet right now. Not every student is doing this, but the option exists, and the students who want to find it will find it on TikTok or YouTube Shorts. They're not going to see videos about responsible AI use in their feed. They're going to see "how to make your Google Doc look like you wrote it yourself." That's one of the reasons I'm a big believer that educators have to be the ones introducing these tools to students. You can't win the blocking game. Your administration decides what goes into your Google Apps for Ed account, and they might listen to your input, but probably not. Or a committee takes three months to decide. In the meantime, you're stuck with whatever's available, and you have to make it work. ## Three layers of AI tools I would actually try Lindy asked me to walk through the tools, and I think a layered approach makes the most sense. The first layer is the agents built into ChatGPT and Claude. These have been around for over a year. In ChatGPT, you hit the tools button, turn on agents, and ask it what it can do for you. If you've been using ChatGPT regularly with the memory setting on, it already knows your context. It knows you teach high school social studies or that you run a department. You don't have to give it background. Just ask it for customized suggestions and go from there. The second layer is Gemini built into Chrome. If you use Chrome, and most K-12 educators do, there's now a Gemini button in the top right corner. You click it, and it can see whatever tabs you have open. No more downloading a Google Doc, uploading it to ChatGPT, and copy-pasting links. Gemini already has the context of your open tabs. For teachers, that removes a few annoying steps. For students, it removes the same friction that was supposed to slow them down enough to think. The third layer is standalone agentic browsers. This is where Comet, Atlas, and Dia live. What Google and Claude did was put extensions into Chrome and let those control the browser. Perplexity took a different approach with Comet. They forked Chrome's architecture, kept your Chrome extensions working, and rebuilt the whole thing with the agent baked in from the start. I tested Comet with the latest Opus 4.6 model, and it ran a task autonomously for about 30 minutes, taking somewhere between 175 and 190 actions as it scrolled, clicked, typed, and navigated. It's not faster than doing it yourself. Maybe half the speed. The useful part is that you set the task, let it run in a tab group, and go do something else. Twenty minutes later, it's done or it tells you where it got stuck. Atlas does the same thing but uses your ChatGPT account, so it's better if you're already invested in that ecosystem. Dia browser is another one worth looking at. Same company that built Arc browser, which is still probably my favorite browser of all time. They made an AI-native version of it. Free tier available, but the good models require a subscription. ## Codex can help with folder work too When OpenAI released Codex in early February, I spent an entire weekend in it, not because I had to, but because it clicked. A lot of the vibe coding and AI coding apps haven't felt relevant to teachers. What would an English teacher do with a coding tool? Codex changes that. It's ChatGPT's model built specifically for code, running in a standalone app with local file access. You don't need a terminal. You don't need to know what a terminal is. Here's what that means in practice. I do some analytics reporting. I can download CSV files, put them in a local folder, open Codex, and say "go analyze everything in this folder." Before, I'd be uploading and downloading files through a browser tab. Now it just reads what's on my machine. I took a website that originally took me 10 to 12 hours to build over multiple revisions. I told Codex to redo the whole thing, and it rebuilt the site with a new design, new content, and a different domain in about thirty minutes. Then I thought, okay, it's good at design. What's a design task that a non-technical person would actually do? So I had it make my LinkedIn banner. It produced something better than what I could do in Canva or Figma, in about a minute. And it saved directly to a folder on my machine without exporting artifacts or downloading PDFs. I just said "put it in this folder and call it designs." OpenAI publicly said they used their newest model to build Codex. The tool was already being used to build the next version of itself. ## OpenClaw is an empty house without utilities The saga of ClaudeBot to MoltBook to OpenClaw made for good social media content, but the actual tool is more limited than the hype suggests. OpenClaw doesn't do anything new. You could have duct-taped your own version together with Claude Code or other tools if you were technical enough. What's different is the interface. It made that kind of setup accessible to regular people. You can message your agent through WhatsApp, and it can search the internet or reach into your cloud files and calendar. But the app is free in the same way an empty house is free. You still need to connect it to the power grid. It needs an AI brain, a ChatGPT or Claude API key, and those cost money to run. The Best Buy in San Francisco ran out of Mac Minis because people were buying dedicated machines to keep OpenClaw running 24/7. Most of those people are going to figure out pretty quickly that the tool is free but the services cost real money. My honest take: you'll spend more time trying to make OpenClaw work than you'll find useful things for it to do. The one real advantage is making AI accessible to people in your life who aren't technical. I set up a version where my wife can text a question about our family calendar, where's the next wrestling meet, what time is student council in the morning, and get an answer in plain language through Telegram. That one job is the useful part. And whatever you do, run it on a separate device. Don't give it access to your work computer, your school email, or anything with sensitive information. Think of it this way: you would not give a student teacher your bank account password and your personal laptop 24 hours a day. Don't give that to an agent either. ## Where this leaves educators Lindy made a comment near the end of our conversation that I keep coming back to. Formal education isn't going to stay relevant if it doesn't adapt. I agree with that, and I'd add a qualifier: adaptation means understanding what these tools can do, not adopting every one of them, so you're not blindsided when a packed room full of educators encounters the word "agentic" for the first time. The students are going to find these tools whether we introduce them or not. They already have. We can show them what responsible use looks like, or we can leave that job to their TikTok algorithm. I don't have a perfect answer. But I'd rather be the teacher who walked down the hall and let my students watch the browser work on its own, came back, and had an honest conversation about what that means, than the one who pretended it wasn't happening. ### Where to find Lindy Hockenbary - [Make EdTech 100 Podcast](https://makeedtech100.transistor.fm/episodes/education-in-the-ai-agent-era-with-aaron-makelky) - [LinkedIn](https://www.linkedin.com/in/lindyhockenbary/) - [Instagram](https://www.instagram.com/lindyhoc/) - [YouTube](https://www.youtube.com/@LindyHoc) ## Related pages - [All AI notes](/blog) - [Tools](/tools) - [About Aaron](/about) ## Sources and references - [Descript](https://www.descript.com/) - [Perplexity](https://www.perplexity.ai/)