Atlas under the Microscope
Agentic Browsing and the Classroom
Smart Teaching Evolved - Issue #13
Monday, October 27, 2025
Welcome to Smart Teaching Evolved—your Monday AI briefing for education professionals. We’re building AI literacy one practical tool at a time, cutting through the hype to find what actually serves teaching and learning. This time we are addressing some emerging technology that has lots of promise (but lots of peril as well).
Intro to AI: What’s an Agentic Browser?
Remember when “browsing the web” meant you had to click every link, fill every form, and copy-paste everything yourself? Agentic browsers are changing that game.
An agentic browser is AI that can actually use websites for you. Not just answer questions about the web—it can navigate pages, click buttons, fill out forms, scroll through content, and gather information across multiple sites. It is like a research assistant who can work through your browser while you focus on higher-level thinking.
Traditional AI (like ChatGPT) has knowledge up to a certain date. Search-enabled AI can look up current information. But an agentic browser can interact with websites the way a human would—logging into systems, comparing data across tabs, and completing multi-step web tasks.
The real power? You can delegate entire research workflows: “Compare the science standards for photosynthesis across these five state education websites and create a summary table.” The agent browses each site, extracts relevant info, and compiles results—work that would take you an hour happens in minutes.
After trying ChatGPT’s Atlas browser, released a few weeks ago, it looks amazing, but then I started thinking about the problems this will pose. I had considered what this might mean beforehand, but now I realize the severe security nightmare this can be. If your browser can log into sites as you, those sites can just as easily take your other information that the browser might have on you. It can be susceptible to “prompt injection” flaws that repurpose your AI.
But let’s be clear about limitations: These tools work best for public information gathering. They’re not (yet) great at subjective judgment calls, and they can make mistakes when websites have unusual layouts. Always verify critical information, especially when it’ll be used in instruction.
Innovative Use: The Research Shortcut That Actually Works
Let’s look at how a high school history teacher used an agentic browser to transform a time-consuming task into a 10-minute win. This is actually a set of instructions that an AI can follow. Some researchers could consider this just agentic behavior, but for us it shows the power of an LLM.
The Challenge: Sarah wanted to find primary source materials about women’s suffrage from different perspectives (suffragists, anti-suffragists, newspaper coverage) for her 11th grade U.S. History class. She wanted sources at appropriate reading levels with proper citations. Normally, this meant hours of browsing archives, evaluating reading complexity, and copying citation information.
The Agentic Approach: Sarah gave this prompt to Claude with computer use enabled:
“Search the Library of Congress digital archives for primary sources about women’s suffrage from 1910-1920. Find 3 documents from suffragists, 2 from anti-suffragists, and 2 newspaper articles. For each, provide: the title, date, a 2-sentence summary, the direct link, and reading level estimate. Focus on documents that would work for 11th graders.”
The agentic browser:
Navigated to LOC digital collections
Searched relevant terms
Opened individual documents
Extracted key information
Evaluated reading complexity based on sentence structure and vocabulary
Compiled everything into an organized list with working links
The Result: In 8 minutes, Sarah had a curated list of seven primary sources with context and citations—work that previously took her 2-3 hours. More importantly, she spent her saved time designing the actual lesson and discussion questions, not gathering materials.
The Key Insight: Agentic browsers excel at structured information gathering across multiple web sources. They’re not replacing your pedagogical expertise—they’re eliminating the grunt work so you can focus on instructional design.
AI News Alert: Computer Use Goes Mainstream—And Teachers Are Already Leading Adoption
Anthropic released its “Computer Use” capability this October, allowing Claude to control computers like a human would—moving the mouse, typing, clicking, and navigating websites. While still in beta, it represents a major shift in how AI can assist with practical tasks.
But here’s what caught my attention: educators are already way ahead of the curve on AI adoption, with 84% of U.S. teachers now actively using AI tools in their work. Even more striking, 60% of teachers who use AI report saving up to 6 hours per week—that’s nearly a full workday back in your schedule.
The Research Reality Check:
Recent data paints a complex but ultimately encouraging picture of AI in education:
65% of educators report feeling more passionate about teaching because of AI, reversing the narrative that technology diminishes the profession
Teacher familiarity with ChatGPT jumped from 55% to 79% in just one year, showing rapid usage
Teachers primarily use AI for creating worksheets, modifying materials for different student needs, and assessment development—saving an average of 5 hours weekly on these tasks alone
But there’s a significant gap: 71% of K-12 teachers have received no professional training on AI, which means most educators are self-teaching—spending time outside work to build these skills. It also might show a limit on self-teaching by reaching an effective understanding of AI, but not pushing it’s limits.
What This Means for Agentic Browsers:
The arrival of computer-use AI comes at exactly the right moment. Teachers have already demonstrated they’ll adopt AI tools that genuinely save time on administrative work. Agentic browsers take this to the next level:
Administrative relief: Routine web-based tasks (finding resources, comparing options, gathering data) can be automated
Research acceleration: Literature reviews, policy comparisons, and resource gathering become dramatically faster
Accessibility gains: Faculty who struggle with complex web navigation have a powerful new support tool
The catch? These tools require thoughtful prompting and oversight. They work best with clear, specific instructions and tasks that involve public information. They’re not yet reliable for high-stakes decisions or tasks requiring nuanced judgment.
Other players in this space: OpenAI is testing similar capabilities, and browser-specific agents like Browserbase are emerging. The technology is maturing rapidly, but education-specific applications are still being developed.
Bottom line: Teachers are already using AI primarily to support students with learning differences and to accelerate routine tasks. Agentic browsers are the natural next step—moving from generating content to gathering and organizing it. Start exploring now with low-stakes tasks to build your understanding before they become standard tools.
The major problem with agentic browsing is that we are seeing the browsers able to complete online courses autonomously. AI can watch videos, produce summaries, participate in class discussions – all without a human intervening. This of course is a significant issue that needs to be considered with agents. Ethics are key to incorporate in all AI considerations.
Try This Monday: Your First Agentic Browser Task
Let’s start with something practical and low-risk that demonstrates the power without overwhelming complexity.
The Task: Use an agentic browser to research professional development opportunities in your area.
What You’ll Need:
Access to Claude.ai (Pro or Team plan with computer use enabled)
15 minutes
A clear research goal
Step-by-Step:
Frame Your Request Clearly:
Bad: “Find PD for me”
Good: “Search for professional development workshops on AI in education within 100 miles of [your city] happening between now and December 31st. For each opportunity, provide: title, date, location, cost, and registration link.”
Give the Agent Context: Add: “I’m a [your role] interested in [specific focus]. I prefer [in-person/virtual/hybrid] options.”
Let It Work: The agent will search multiple sites, compile results, and organize them for you.
Review and Verify: Check 2-3 results to ensure accuracy. Agentic browsers are good but not perfect—always verify before committing time or money.
What You’ll Learn:
How to structure requests for best results
What kinds of tasks work well (structured information gathering)
What still needs human oversight (judgment calls about relevance)
Reflection Questions:
How long would this task have taken manually?
What would you do with that saved time?
What other research tasks follow this pattern?
This simple exercise builds your intuition for when agentic browsers add value versus when manual searching is still better. Start here, then scale to more complex research tasks as you gain confidence.
Quick Tip: When NOT to Use Agentic Browsers
Agentic browsers are powerful, but they’re not the right tool for every job. Skip them when:
You need creative exploration: Browsing with open curiosity often leads to unexpected discoveries. Agents are goal-directed, not serendipitous.
The task requires subjective judgment: “Find engaging articles” is too vague. “Find articles about photosynthesis published in 2024 from .edu domains” works better.
You’re working with sensitive information: Don’t use agentic browsers for anything involving student data, passwords, or confidential information.
Speed isn’t the priority: If you’re learning about a new topic, manual browsing helps you develop understanding that automated searching misses.
The sweet spot: Structured, repetitive web research that follows clear patterns. Think gathering, not discovering.
Got questions about agentic browsers or AI in your educational setting? Hit reply—I read every email and will feature subscriber questions in future issues.
📧 Forward this to a colleague who spends too much time on web research
💡 Share your AI experiments—your stories help other educators
🎯 Missed an issue? Find the archive at robvossai.substack.com
Smart Teaching Evolved is published by Dr. Robert Voss
Voss AI Consulting | VossAIConsulting.com | robvoss@vossaiconsulting.com
Helping education professionals and organizations understand and leverage AI for student learning and institutional success.
You’re receiving this because you subscribed at robvossai.substack.com.
P.S. The best agentic browser prompts are the ones that save you time on tasks you’d procrastinate anyway. What’s your least favorite web research task? Reply and tell me—I might feature solutions in a future issue.
This piece really made me think: what if it personalize student learning?