Every Student Will Have AI Agents Sooner Than Later: The Question Is Whether Schools Will Be Ready

The shift is not coming. It is already underway and the pace is about to accelerate beyond what most educators are prepared for.

A few weeks ago, I published a piece about how I built a personalized study system using Claude and the principles of the science of learning : retrieval practice, spaced practice, interleaving, and formative assessment and feedback with immediate correction. I designed it deliberately, uploaded my materials, built flashcard banks, mock assessments, blocked study activities, and a mastery tracker I could select after each study session I had over the course of the week. It worked. It worked well enough that I walked into a high-stakes performance feeling genuinely fluent rather than passingly familiar.

Here is the part I want teachers and school leaders to sit with: I had to build all of that myself. I had to know which cognitive science principles to apply. I had to prompt for the right artifacts, sequence my own study sessions, and make deliberate decisions about spacing and interleaving. The intelligence was mine. Claude was the tool I put it all into motion.

That is not how it will work for the next generation of students.

What an Agent Is and Why It Matters

An AI model as a traditional chatbot answers questions. An AI agent acts on the world. It is AI outside the box.

The distinction sounds subtle. It is not. It is enormous. A traditional model as a chatbot responds to what you give it. An agent pursues goals, monitors progress, makes decisions, triggers other agents, accesses memory across time, and adjusts its approach based on what it observes about you. Agents like Hermes, and OpenClaw represent early iterations of this architecture. Claude Code and OpenAI’s Codex also fall into this bucket to a degree. They are systems that can operate autonomously, hand off tasks to specialized to one or swarms of subagents, manage persistent memory, and even transact on your behalf. These are not chatbots with more features. They are different in what they can do and who is directing the work.

Now apply that architecture to a student.

An agent built for learning would not wait to be asked for a flashcard set. It would monitor a student’s performance across assignments and assessments, identify the specific concepts where retrieval is failing, automatically generate targeted practice using spaced repetition schedules calibrated to that student’s forgetting curve, and surface the right content at the right time; whether that student opens a study tool or not. It would know the student’s prior knowledge, their learning history, their upcoming assessments, and their current mastery level. It would build what I built manually, automatically, for every student, every day, without requiring the student to know anything about cognitive science.

That is not a hypothetical. That is the logical and near-inevitable endpoint of the capabilities already deployed in systems like Hermes and OpenClaw today that I have experimented with over the last six months. It is here.

The Science Makes the Case

The reason this matters so much in education specifically is that the science of learning is clear about what works and equally clear about how rarely we do it.

Retrieval practice outperforms re-reading by a significant margin. Spaced practice outperforms massed study. Interleaving outperforms blocked practice. Formative assessment with immediate, specific feedback is one of the highest-impact interventions in all of education research. We have known these things for decades. Bjork, Roediger, Karpicke, Ebbinghaus, and Sweller gave us the empirical foundation. What we have never had is the infrastructure to deliver it to each and every student, personalized to every student, in every classroom, in every subject, at every grade level.

AI agents are that infrastructure.

A well-designed agent does not generate content. It sequences it. It spaces it. It interleaves old material with new. It administers low-stakes retrieval checks at intervals that match the research on optimal spacing. It provides feedback that is specific enough to close the gap between what a student produced and what a strong response would have included. This is not what a teacher with 34 students and 47 minutes of instructional time can do manually. It is exactly what an agent operating in the background; persistent, patient, and data-rich. It can do continuously without ever stopping.

What Hermes, OpenClaw, and Their Successors Will Do

Current systems like Hermes and OpenClaw already demonstrate the architecture that future student-facing agents will build on. They can operate as orchestrators , directing specialized subagents to handle discrete tasks while maintaining a unified memory and goal structure. They can exchange information across sessions, trigger actions based on observed patterns, and function across platforms and tools without requiring a human to manually connect each step.

For students and teachers, this translates to agents that will do the following and much more. Paired with super power AI like Mythos and beyond, there are exponential opportunities for the agent to assist and amplify learning.

Know everything academically relevant about the student. Not grades. Mastery patterns, misconception histories, engagement data, prior knowledge gaps, and performance under different instructional conditions.

Build and rebuild learning sequences automatically. An agent reviewing a student’s recent quiz performance does not produce a summary. It generates a targeted practice set, schedules it for the optimal review window, and delivers it through whatever interface the student uses most.

Support lesson design for teachers. A teacher describing an upcoming unit would receive a mapped sequence aligned to cognitive load principles, formative checkpoints built in, and student-specific scaffolds generated from each learner’s profile.

Operate continuously, when prompted. This is the most significant departure from current AI tools. An agent does not wait for a student to open an app. It works between sessions, prepares materials before the student arrives, and flags patterns for teachers before they become failures.

Learn and improve over time. Each interaction refines the agent’s model of that student. The longer the relationship, the more precise the support. This is personalization at a level no human teacher could sustain across a full roster.

The Uncomfortable Question for Teachers and School/District Leaders

I want to be direct here, because the edtech conversation about AI agents tends toward either breathless optimism or anxious resistance and neither is useful.

These agents are coming. The economic incentives are too strong, the technical capabilities are advancing too quickly, and the scale of educational need is too large for this not to happen. The question is not whether every student will have a self-improving AI agent working alongside them by the end of this decade. The question is whether educators will have shaped what those agents do, how they operate within schools, and what values and principles guide their design.

Teachers who understand the science of learning will be better positioned to evaluate whether an agent is supporting learning or generating the appearance of productivity. Administrators who understand agent architecture will ask better questions about data privacy, equity of access, and instructional coherence. Curriculum designers who have studied cognitive load theory will recognize when an agent’s scaffolding is reducing desirable difficulty in ways that undermine long-term retention.

The professionals who matter most in shaping how these tools work in schools are educators , but only if educators engage with what these systems are before someone else defines that for them.

Putting It All Together

AI agents are not a future possibility. They are an emerging reality that will reach every classroom within the next few years. The districts that prepare now, by building teacher fluency, asking the hard policy questions, and grounding their approach in the science of learning, will be the ones whose students benefit most.

What to Do Right Now

You do not have to wait for a fully autonomous agent to start building the fluency you will need.

Start by doing what I described in my previous piece. Use Claude, Codex, Gemini, or any capable AI tool to build study systems grounded in retrieval practice rather than passive review. Prompt for interleaved practice, spaced repetition structures, and formative assessments with scoring and feedback. Use these tools with students deliberately, not because they are novel but because the cognitive science says they work.

Learn what agents are and how they differ from standard AI models. Follow the development of systems like Hermes and OpenClaw not as curiosities but as early signals of a design trajectory.

However, start asking the questions your district, your school, and your students will need answered before the agents arrive: Who owns the student data? How does an agent’s personalization interact with IEP requirements? What happens to teacher judgment when an agent is already making instructional decisions? What does equitable access to AI agent support require?

Ultimately, the agents are coming and they are already here. The school and district leaders along with teachers who have already been asking the hard questions will be the ones whose students benefit most when they arrive.

---