Today's Agenda
Hello fellow humans! Today, there is big news around agents, and for most people, it may not be clear what that means, but the AI powerhouse companies are trying to reshape our entire information ecosystem, and the pace of change is breathtaking. OpenAI and Nvidia have both declared that the 2025 is the year of the agent.
Agents are pieces of software that use AI for you to do work on your behalf. Imagine JARVIS from the Marvel Cinematic Universe as an agent who can call upon a team of other agents and AI models to perform several tasks at the same time, in parallel, to accomplish work for you. Theoretically, agents can use AI models to research topics, craft emails, update documents, schedule activities for you, set to-do items, review email chains to create meeting agendas, and much more.
As you’ll see, the results so far are mixed. But Microsoft is bullish on the outlook for agent capabilities, and the breakneck pace of AI change is a strong signal that we may soon see AI agents that can be trusted to execute work for us.
News
A Company Without Humans?
Researchers at Carnegie Mellon University ran an experiment running a fully autonomous AI Agent organization called TheAgentCompany. AI agents filled a number of roles such as analysts, software engineers, an AI-driven HR department, and even the Chief Technical Officer.
Business Insider reported that it did not go well. The results were expensive, inefficient, and had an abysmal success rate.
But increasingly, achievements that experts expect in 1-2 years are now being achieved in six months. So this study is just one data point in time, revealing weaknesses that could be fixed faster than you think. That should also be taken with a grain of salt as OpenAI’s ChatGPT 5 has been delayed “a couple of months.”
The Achilles Heel of AI
Elon Musk is showing us putting his thumb on the scale as he publicly pushes back against his own Grok chatbot. The article tells a story of an X-formerly-Twitter user who posted a conspiracy-laden article that got fact-checked by the Grok chatbot, and Musk was not happy about it.
There is also a slow-growing body of evidence that Musk is leveraging his xAI company for his own political agenda. And while these are small, limited examples of AI being directed by humans, these are a warning. Chatbots can sound very confident and even persuasive, and we need to be appropriately skeptical of what generative AI serves us. Our best defense, whether we use AI or encounter people informed by AI, is to develop habits of mind, leveraging our critical thinking and literacy skills to sort facts from manipulations and hallucinations.
The utility of language models changes if they are vulnerable to manipulation by their operators. Optimization is a function of the objective, and when the product is optimized for owners and not users, the users become the product. As the saying goes, “If something is free, you’re not the customer; you’re the product.”
Microsoft is Driving Agents
Hot on the heels of Microsoft’s report on Frontier Firms integrating AI agents into the workforce, the software giant at Microsoft Build 2025 announced major updates to Microsoft 365 Copilot. Microsoft expects this to be a game-changer — imagine you’re now a team lead for a group of agents that can autonomously handle work for you.
“Update the slide deck based on the feedback from the VP’s email.”
“Put together the onboarding packet for this new employee and set up access and accounts.”
“Review these historical documents and make three strategy recommendations with a pitch deck for each.”
“Review my inbox daily, update my To-Do list based the most important conversations, and propose a schedule of calls to discuss open questions and write an agenda for each call.”
You’re no longer bouncing back and forth between tasks in Excel, Word, Outlook, and PowerPoint, but instead directing a team.
If you’re using it just for efficiency, I think you’re under-utilizing it. That is our big learning from observing people who use it really great versus folks who have AI access, who have CoPilot, but they're using it still in… ways that are incremental.
She lays out a vision for work where your agents can review your past interactions and make recommendations for meeting agenda items, can assemble an onboarding packet, create accounts and provide access for new employees, generate updated marketing materials for a new campaign or event, or complete documentation for regulatory compliance.
The Neuron article includes compelling use case results, so I strongly recommend reading the original piece.
The big takeaway for me is that agents can be powerful tools to automate work, but the most successful people will be those who have strong:
Communication skills for crafting those expectations
Collaboration skills for architecting how those agents work together
Critical thinking skills for active monitoring of the ongoing work to ensure that the agents are choosing the right actions, doing it correctly, efficiently, and effectively.
All of this requires a high level of trust in the AI systems that Microsoft is building, and that may require a long period of skepticism, scrutiny, and fine-tuning using these critical cognitive skills.
Skill Stack
Five Education Strategies for the AI Era
Here are five research-backed instructor strategies to avoid the traps of AI cheating and teach the critical cognitive skills that students will need as they enter a world that will be dominated by AI technologies.
Conversation is the key to unlocking both human cognitive powers and the power of AI. Just like having a conversation with a chatbot can help you clarify your own ideas and thinking, having conversations with students before, during, and after the learning work has some powerful force-multiplying effects. Here are five strategies you can leverage to help your students develop those cognitive skills:
Oral Exams and Defenses > Written Assessments
Once upon a time, exams were oral, and students had to present defenses of their work to instructors, peers, and assessment boards. This fell out of practice because it does not scale well. Classes were smaller, fewer children were expected to be educated, institutions were staffed for the practice, and writing was closer to final publication than a way to sketch and organize thoughts like it is today. Today, the use case is that oral presentations and defenses inherently limit AI reliance, as students must articulate and defend ideas spontaneously.
Oral exams and presentations require students to demonstrate real-time critical thinking (analysis, synthesis, evaluation), which written papers may obscure due to editing or external assistance.
Follow-up Q&A sessions allow instructors to probe understanding depth, reducing opportunities for AI-generated content to go undetected.
Students in oral assessments showed higher material retention and reduced plagiarism rates compared to written exams.
Leveraging whiteboards or online drawing tools can make assessments multimodal and more interactive.
Instructors could even create and use a custom GPT AI interviewer as part of the process.
Scaffolded Presentations and Collaborative Learning
When instructors actively coach their students through a process, reviewing and giving feedback on student work progressively instead of only providing feedback on the final deliverable, the instructor can see the entire thinking process that went into the work. In these direct, human social interactions, it is harder for the student to inappropriately use AI. The instructor may even allow AI within certain guidelines to evaluate the inputs and outputs together. Multi-stage assessments with oral components can disrupt students delegating work to AI by requiring iterative, personalized engagement with the instructor.
Collaborative research presentations and group debates force students to internalize material, making AI-generated content harder to integrate undetected.
Collaborative learning teaches the skills for using AI more critically and effectively when they do use AI.
Scaffolded assignments (e.g., incremental drafts + peer feedback) paired with final presentations reduce "last-minute AI outsourcing."
Curriculum Co-Creation and Self-Directed Learning
We found several studies showing that students engaged in the curriculum development process enhanced executive functioning (planning, self-monitoring) and creative thinking through iterative curriculum design tasks. One study focused on Design Thinking as part of a collaborative curriculum co-creation process in determining assignment topics, developing learning objectives, and contributing to the design of grading assessments. The study found:
Students who co-designed curriculum components (assignments, learning objectives, grading criteria) demonstrated heightened engagement and metacognitive awareness of their learning processes.
Participation fostered ownership of educational outcomes, with students reporting better performance and deeper understanding of project goals.
The design-thinking approach strengthened problem-solving skills and collaborative abilities.
Digital Literacy and Presentation Skill Gaps
When students teach a subject as part of their learning process, they quickly find the gaps in their information and seek out strategies and tactics to fill those gaps. That’s the core insight behind the Feynman Technique, and this study on presentation skills in nursing schools reaches a similar conclusion. This study suggests techniques to aid the process of discovering those gaps by identifying how the discrepancies between written and oral performance raise red flags for AI misuse.
Structured presentation rubrics (e.g., organization, delivery, audience engagement) reveal gaps in applied understanding that written papers may mask.
Students with weak presentation skills often overcompensate with AI-polished written work, which oral assessments expose.
Analyzing the differences between the oral presentation and the written work highlights the students’ knowledge and skill gaps.
Ironically, instructors skilled in using AI can achieve scale and conduct multi-modal curriculum design, active student engagement, and assessments. Just as Microsoft envisions employees having a team of agents to accomplish work for them, teachers can have a team of agents that can transcribe, parse, and analyze oral exams, or even prompt the student with questions during an oral presentation. Another agent could compare the spoken presentation to the written work and highlight discrepancies to the instructor.
Of course, all of this hinges on a level of trust for AI models that we have not fully achieved yet.
Radical Candor
We are at the dawn of this radical transformation of humans that by its very nature is a truly complex and emergent innovation. Nobody on earth can predict what’s gonna happen. We’re on the event horizon of something… This is an uncontrolled experiment in which all of humanity is downstream.