[{"data":1,"prerenderedAt":252},["ShallowReactive",2],{"blog-agentic-learning-and-the-courses-it-wont-replace":3},{"id":4,"title":5,"author":6,"body":7,"date":236,"description":237,"extension":238,"image":239,"meta":240,"navigation":241,"path":242,"seo":243,"stem":244,"tags":245,"__hash__":251},"blog\u002Fblog\u002Fagentic-learning-and-the-courses-it-wont-replace.md","Agentic learning, and the courses it will not replace","Peter",{"type":8,"value":9,"toc":224},"minimark",[10,19,22,27,30,33,36,40,43,46,49,53,60,66,72,78,84,90,93,97,104,115,122,125,135,139,142,145,149,156,159,162,180,183,187,190,193,196,200,203,206,209,212],[11,12,13,14,18],"p",{},"A learner types their goal into a chat window. An AI agent takes it from there — diagnosing what they already know, generating an explanation, posing a question, watching the answer, choosing what comes next. No predefined curriculum. No built lessons. The course as an artifact does not exist; the agent ",[15,16,17],"em",{},"is"," the course.",[11,20,21],{},"This kind of agentic learning is real now, for some subjects, with some learners. Whether it should replace structured courses, for everyone and everything, is a more interesting question than the marketing usually allows. The honest answer is that the early evidence is mixed, that the structural problems are not solved, and that the most useful frame is not \"agentic versus structured\" but \"which parts of which learning experience should be each.\"",[23,24,26],"h2",{"id":25},"what-agentic-learning-actually-means","What \"agentic learning\" actually means",[11,28,29],{},"The term gets used loosely, so worth pinning down. By agentic learning I mean a system where an AI agent owns the whole loop: assessing the learner, deciding what to teach next, generating the content, checking whether it stuck, adapting from there. Nothing pre-authored. Nothing version-controlled. The interaction is the entire experience.",[11,31,32],{},"That is meaningfully different from two adjacent things LearnBuilder already does. AI-assisted authoring means an AI helps a human build a course you then deliver — the artifact remains. An AI tutor inside a structured course means a chat helper lives next to fixed lessons — the artifact frames the interaction. Both are useful and both exist today. Pure agentic learning, in the strict sense, is the case where there is no artifact at all.",[11,34,35],{},"This piece is about that strict case.",[23,37,39],{"id":38},"what-an-agent-can-genuinely-do-right-now","What an agent can genuinely do right now",[11,41,42],{},"Conversational explanation that adapts to questions, with patience that no human tutor sustains for free. Practice problem generation in mathematics, language learning, basic programming, and some scientific reasoning. Reasonable inference of a learner's level from a few exchanges. 24\u002F7 availability at near-zero marginal cost per learner. Tutoring quality that is genuinely useful for high-resource domains and major languages — the things the underlying model has seen enough of.",[11,44,45],{},"The most visible production examples are Khan Academy's Khanmigo, which wraps GPT-4 in a constrained tutoring interface, and Duolingo Max, which uses LLMs for conversational language practice and explanation. The largest user population, almost certainly, is informal — students using ChatGPT or Claude as a tutor with no product wrapper around it at all.",[11,47,48],{},"For a curious adult exploring a topic, a child supplementing schoolwork, a developer picking up a library on a Saturday afternoon, this works. The personalisation, the patience, and the price point are all real improvements over the realistic alternative for most learners, which is no tutor at all.",[23,50,52],{"id":51},"what-it-still-cannot-do-and-why-each-gap-matters","What it still cannot do — and why each gap matters",[11,54,55,59],{},[56,57,58],"strong",{},"Reliable assessment of mastery."," The agent can tell whether a learner produced a correct-looking answer in chat. It cannot reliably tell whether the underlying knowledge will transfer to a different context two weeks later. Engagement and learning are not the same thing, and current systems optimise the former by default.",[11,61,62,65],{},[56,63,64],{},"Coherent multi-week curriculum design."," A model can sequence three explanations in a row. Whether the seventh week of a course should revisit week two's concepts, in what form, with what spacing, with which retrieval prompts, is the kind of structural judgement that benefits from intentional design rather than turn-by-turn improvisation.",[11,67,68,71],{},[56,69,70],{},"Verified factual accuracy in specialised or recent domains."," Hallucination in general knowledge has been reduced substantially. In compliance, medicine, law, internal company knowledge, and recently updated regulations, the failure modes are still material and still confident. Retrieval helps; it does not eliminate the problem.",[11,73,74,77],{},[56,75,76],{},"Production-quality interactive media on demand."," Branching scenarios, accurate diagrams, simulations, well-edited video, drag-and-drop activities — these can be generated, but the gap between \"generated\" and \"well-designed\" is large in 2026. For most domains, the well-designed version still takes a human deciding.",[11,79,80,83],{},[56,81,82],{},"An auditable trail of what was taught."," Compliance is the obvious case, but professional certification, regulated industries, accreditation, and most workplace L&D need a thing to point to that says: this is the content, this is the version, these are the outcomes, here is the evidence learners met them. A pure agentic system, by design, does not produce that artifact.",[11,85,86,89],{},[56,87,88],{},"Cohort dynamics."," Group discussion, shared deadlines, peer-explanation effects, the social motivation of a course running in real time with other humans on the same schedule. An agent can produce content. It does not produce a class.",[11,91,92],{},"Each of these gaps is structural, not a temporary engineering limitation that closes in eighteen months. Some will improve materially. Others — particularly the artifact and the cohort — are not really problems the agentic frame is trying to solve.",[23,94,96],{"id":95},"what-the-research-actually-says","What the research actually says",[11,98,99,100,103],{},"The headline number people quote in favour of AI tutoring is Bloom's \"2 sigma\" finding from 1984 — that one-on-one human tutoring produced learning gains roughly two standard deviations above a conventional classroom. The implied argument is that AI agents can replicate that economically. The trouble is that the original study is about ",[15,101,102],{},"human"," tutors, and the empirical record for LLM tutors is genuinely mixed.",[11,105,106,107,110,111,114],{},"The single most useful study to know about is Bastani et al. (2024) at Wharton, ",[15,108,109],{},"Generative AI Can Harm Learning",". They ran a controlled experiment in a high-school mathematics context. Students with unrestricted access to GPT-4 during practice scored ",[15,112,113],{},"worse"," on a follow-up exam than students without GPT-4 access at all. A constrained \"tutor mode\" version — which guided students rather than solving for them — performed better. The mechanism is the part that matters: a too-helpful agent removes the productive struggle that builds durable knowledge. Practice without resistance does not generate retention.",[11,116,117,118,121],{},"This connects to long-established cognitive science. Robert Bjork's work on \"desirable difficulties\" predicts that frictionless explanations, given before a learner has wrestled with a problem, reduce retention. John Sweller's cognitive load theory predicts something compatible — that ",[15,119,120],{},"too much"," scaffolding can be as harmful as too little. Agentic systems optimised for user satisfaction will lean toward frictionless, because that is what learners ask for in the moment. That is exactly the failure mode the research warns about.",[11,123,124],{},"On the other side of the ledger, there are early reports — not yet a robust RCT base — that constrained AI tutors with explicit pedagogical guardrails (Khanmigo is the example most often cited) modestly help with engagement and self-paced practice. Ethan Mollick and colleagues at Wharton have written about AI co-tutoring in classroom contexts producing reasonable results when the agent is configured to coach rather than solve.",[11,126,127,128,131,132,134],{},"The honest summary: the technology is improving fast, the early research is small-N and short-horizon, and the field has not yet shown durable, transferable learning gains from pure agentic systems comparable to what Bloom's 2-sigma claim promises. If you read one paper before adopting agentic learning at scale, make it the Bastani study. The finding that an ",[15,129,130],{},"unguided"," AI tutor can produce ",[15,133,113],{}," outcomes than no tutor at all is the thing most marketing in this space quietly ignores.",[23,136,138],{"id":137},"what-this-article-is-probably-getting-wrong","What this article is probably getting wrong",[11,140,141],{},"I am cautious because the current evidence is mixed and the structural problems — assessment, traceability, productive struggle — are not solved. But agentic systems are improving quickly. Bastani used GPT-4 in 2023. A 2026 system with better retrieval, stronger guardrails, real assessment hooks, and explicit pedagogical scaffolding could produce a very different result. A future where pure agentic learning genuinely matches structured courses for a much broader range of contexts is plausible, and it could arrive faster than this piece implies.",[11,143,144],{},"Treat the \"not ready to replace\" claim as load-bearing on today's evidence. If the evidence shifts, the claim shifts. The structural points about the artifact and the cohort survive technological progress; the empirical points about learning outcomes do not necessarily.",[23,146,148],{"id":147},"why-structured-courses-survive-where-the-artifact-matters","Why structured courses survive where the artifact matters",[11,150,151,152,155],{},"A course is a ",[15,153,154],{},"durable artifact",". It can be versioned, reviewed, signed off by a compliance officer or subject expert, branded, translated, deployed across an LMS, and completed by a learner with a record that completion happened. These are not nice-to-have features. For workplace training, professional certification, regulated industries, and any context where what was taught matters to someone other than the learner, the artifact is the point.",[11,157,158],{},"Agentic learning, by design, does not produce that artifact. The conversation happened. The learner improved, or did not. There is no canonical statement of \"this is what the course taught\" that an auditor or accreditor can examine.",[11,160,161],{},"So the structural argument is straightforward: built courses survive wherever the artifact matters. That is most workplace L&D, most compliance, most certification, most regulated content, most cohort programmes, and most situations where a stakeholder downstream of the learner has a reason to care what was actually covered.",[11,163,164,165,168,169,174,175,179],{},"What does ",[15,166,167],{},"not"," survive untouched is the assumption that a structured course must be inert. LearnBuilder is already moving toward courses that have agentic capabilities inside them — AI course generation as a drafting tool (",[170,171,173],"a",{"href":172},"\u002Fblog\u002Fai-assisted-hand-built-or-both-learnbuilder","covered in detail in this earlier post","), AI tutoring inside lessons, AI-generated practice variants, AI-translated subtitles across sixty languages (",[170,176,178],{"href":177},"\u002Fblog\u002Flearnbuilder-april-2026-release","from the April release","), AI-generated interactive blocks. The structural unit is the course; the texture inside it is increasingly agentic.",[11,181,182],{},"The bet is hybrid: built courses with rich agentic features inside, not pure agentic loops replacing the course as a concept. Replacement everywhere is unlikely. Replacement in the artifact-mattering contexts is structurally blocked, not just empirically weak.",[23,184,186],{"id":185},"where-pure-agentic-learning-probably-wins","Where pure agentic learning probably wins",[11,188,189],{},"The contexts where the artifact does not matter are real, and large, and growing. A curious adult learning a topic for themselves. A child supplementing schoolwork. A developer picking up a new library on a weekend. A student preparing for a standardised exam. Conversational language practice. Hobbyist skill building. Self-directed exploration of any kind where the learner is the only meaningful stakeholder.",[11,191,192],{},"In those contexts the agentic loop is genuinely powerful, and a built course looks slow and clumsy by comparison. The future of self-directed learning probably is largely agentic, and that is good news — the economics of patient, infinitely-available tutoring at near-zero marginal cost change what is possible for learners who would otherwise have nothing.",[11,194,195],{},"The pure-agentic future is real. It is just not the whole future.",[23,197,199],{"id":198},"the-useful-question-is-which-parts-not-which-side","The useful question is which parts, not which side",[11,201,202],{},"The framing of \"agentic replaces structured\" is a marketing artefact, not a useful design question. The useful question is which parts of a learner's experience should be agentic and which should be structured, and the answer changes with the stakes, the audience, the accountability, and the evidence the learning needs to have produced.",[11,204,205],{},"For self-directed exploration: lean agentic. For certification and compliance: lean structured. For most workplace and educational learning, where neither extreme fits cleanly: build a structured course and put agentic capabilities inside it. The course gives you the artifact, the audit trail, the cohort, the brand, and the version control. The agentic features give you the personalisation, the patience, the on-demand practice, and the explanation tailored to the learner in front of you.",[11,207,208],{},"The research, on current evidence, suggests the gains live in that hybrid space rather than at either pole. The structural arguments suggest the same. Build for both, and revisit the balance as the evidence moves.",[210,211],"hr",{},[11,213,214],{},[15,215,216,217,223],{},"If you want to see hybrid in practice — built courses with AI tutoring, AI-generated practice, and AI-translated subtitles inside — the ",[170,218,222],{"href":219,"rel":220},"https:\u002F\u002Flearnbuilder.org",[221],"nofollow","free trial"," is the place to start.",{"title":225,"searchDepth":226,"depth":226,"links":227},"",2,[228,229,230,231,232,233,234,235],{"id":25,"depth":226,"text":26},{"id":38,"depth":226,"text":39},{"id":51,"depth":226,"text":52},{"id":95,"depth":226,"text":96},{"id":137,"depth":226,"text":138},{"id":147,"depth":226,"text":148},{"id":185,"depth":226,"text":186},{"id":198,"depth":226,"text":199},"2026-05-24","An honest look at agentic learning — AI agents that design and deliver the whole learning experience on the fly. What is genuinely possible today, what early research like Bastani et al.'s 'Generative AI Can Harm Learning' actually shows, and which parts of structured course design no agent has yet replaced.","md","\u002Fblog\u002Fagentic-learning-and-the-courses-it-wont-replace.webp",{},true,"\u002Fblog\u002Fagentic-learning-and-the-courses-it-wont-replace",{"title":5,"description":237},"blog\u002Fagentic-learning-and-the-courses-it-wont-replace",[246,247,248,249,250],"AI","agentic learning","research","instructional design","course authoring","5XuMXcLRdIYmz1JuQCPWEIdAcZVwbyqxXeWAwxed3TA",1779992480040]