There is a difference between an AI that answers a question when asked and an AI that notices you have not opened the learning app in a week, drafts a personalised catch-up plan, delivers the first module, checkshow you did, adapts the next one and messages your teacher when it thinks ahuman should step in.
That gap is what separates the first wave of AI in education(chatbots as reactive assistants) from the second (agentic AI as proactive coaches). The technology to build the second wave is now here. Whether an EdTech platform or a university digital learning team ships something that actually improves learner outcomes depends less on the model choice and more on how the loop, the memory, the guardrails and the economics are designed.
From our work
VE3 has built its cognitive AI capability into a productised platform, PromptX, whose super-agent layer is built on Anthropic's Model Context Protocol and deployed in enterprise settings from finance to higher education. What follows is the pattern we see when EdTech companies and universities start moving from AI as a widget on the page to AI as an actor inside the learning loop.
What "agentic AI" actually means
The word "agent" is overused. In the learningcontext, four characteristics matter.
Autonomy. The system can initiate actions rather thanonly responding to prompts.
Multi-step planning. It can break a goal into stepsand execute them in sequence, calling on different tools or sub-agents asneeded.
Memory. It carries state across sessions, so whathappened last week informs what happens today.
Goal orientation. Every action is measured against anoutcome, typically a learner outcome, that the agent is trying to move toward.
An LLM answering a single question is not an agent. An LLMinside a loop that plans, acts, checks and adapts, with persistent memory ofthe learner, is.
From assistant to coach: the shift that changes learning
Give the same scenario to a first-wave assistant and asecond-wave coach.
A student fails a quiz on quadratic equations.
The assistant does nothing. It waits until thestudent asks a question. If the student does, it explains quadratic equationsone more time. When the student leaves, the assistant forgets everything.
The coach notices the failed quiz within minutes. Itdiagnoses which specific sub-skills the student missed (factorising, applyingthe formula, interpreting the discriminant). It generates a targeted practiceset, delivers it at the student's usual login time, checks the result, andeither advances the student to the next topic or routes them to a human tutorif the confusion persists. It remembers next week.
The move from assistant to coach is not a bigger model. Itis a system that has memory, a goal, and permission to act.
What agentic AI unlocks in the learning loop
Every learning experience runs the same loop: assess,plan, deliver, feedback, adjust. An agentic system can operate that loopautonomously across a whole cohort while still keeping the teacher meaningfullyin charge.
Continuous formative assessment. Every interaction isa data point. The agent updates its estimate of what the learner knows on everyattempt, not on scheduled quizzes.
Personalised curriculum adaptation. The path throughthe material changes for each learner based on demonstrated mastery, without ahuman curriculum designer writing every branch.
Just-in-time remediation. When misconception appears,help arrives inside the same session, not on the next scheduled office hour.
Longitudinal progress tracking. The agent remembersweeks of interactions, not just the current session, which is the preconditionfor genuinely adaptive coaching.
Cross-domain transfer. An agent that has watched alearner struggle with proportional reasoning in maths can flag the sameweakness when it appears in a chemistry problem.
Multi-agent architectures beat monoliths
The instinct is to build one large agent that doeseverything. In production, that pattern is fragile. A multi-agent architecture,with each agent specialised and scoped, is more reliable and easier to govern.
- Tutor agent — delivers content, explains, coaches through worked examples
- Assessment agent — scores work, diagnoses misconceptions, reports gaps
- Planning agent — sequences the next best learning activity for each learner
- Retention agent — spots disengagement signals and triggers nudges or advisor handoffs
- Orchestrator agent — coordinates the others and holds the loop's state
Each agent has a limited job, a verifiable output and atestable behaviour. Guardrails apply per agent. Failures are containable. Thisis how PromptX's MCP-powered super-agents are structured, and it is the pattern that works at scale.
What memory means for a learning agent
Two distinct memory systems matter, and confusing them is acommon early mistake.
Semantic memory of the learner. A structured,queryable profile: mastery state, misconceptions observed, learningpreferences, goals. Lives in a database, updated after every meaningfulinteraction.
Episodic memory of interactions. Every session, everyattempt, every question, every hint given. Typically stored in a vectordatabase so the agent can retrieve semantically similar past interactions.
The best coaches, human and artificial, use both. Theyremember what a learner knows and they remember what happened last time.
There is a third memory that is often overlooked: the memorythe learner has of the agent. Trust and rapport build when the coach remembersthe learner. The single biggest reason first-wave AI tutors did not liftoutcomes is that they forgot the learner between sessions.
The teacher stays in the loop, but the role changes
The teacher becomes an orchestrator, curator andadjudicator. They set the agent's guardrails, review the plans it proposes forthe cohort, intervene at the moments where human judgement matters (motivation,career direction, ethics), and pick up the escalations the agent routes tothem.
Practical human-in-the-loop patterns:
- Teacher-approved curriculum branches for topics with sensitive content
- Escalation triggers when the agent detects distress, confusion beyond its remedy, or safeguarding concerns
- Weekly summary of the agent's decisions per learner, reviewable in ten minutes
- Manual override that always wins, logged for audit
Guardrails, safety and assessment integrity
Agentic AI in learning fails in specific ways that reactiveAI does not.
It can do the work for the learner. A coach thatsolves the problem defeats the point of practising. Explicit rules about whatthe agent will and will not do for the learner, and clear disclosure of AI usein assessed work, are non-negotiable.
It can hallucinate confidently. RAG grounding into anapproved content corpus, with source attribution on every generatedexplanation, is the minimum bar. Anthropic's own Responsible Scaling Policy and theemerging patterns around grounded generation are the starting point.
It can propagate bias in assessment. Regular auditsof agent scoring against human graders are the fix. Silent drift is the risk.
It can leak data. Human-subject data, safeguardinginformation, minors' details need explicit boundaries. The agent should neversee what it does not need to see.
Cost. Multi-step agent reasoning burns tokens.Without a cost-routing layer, an EdTech platform can accidentally 10x itsinference bill overnight.
The economics: when agents earn their keep
Not every learning moment needs an agent. Delivering avideo, marking a multiple-choice question or fetching a definition are jobs asmall model or a lookup can do at a hundredth of the cost.
Agents earn their keep at high-value moments: diagnosticsessions where the agent has to reason across multiple wrong answers to findthe underlying misconception, retention interventions where personalisationdrives measurable re-engagement, and multi-step coaching where a small modelwould not follow the thread.
A cost-optimised model routing layer, like the FinOps layerinside PromptX, sends each request to the smallest capable model rather thandefaulting every query to a frontier one. In VE3's engagements, this has cut AIspend on comparable workloads by roughly a third without impacting outcomequality.
Common failure modes
Anthropomorphising the agent. Treating its outputs asexpert judgement. It is a system with limits, not a colleague.
Loop-of-death. An agent that reasons forever, neveracts, and burns tokens. Time and step budgets are mandatory.
Guardrails retrofitted. Adding safety after launch.Every documented failure has been someone doing this.
Assessment integrity slippage. Agent starts as acoach, quietly becomes a cheat engine. The line has to be defended explicitly.
Data protection retrofitted. Same story. Ineducation, this is not a fine risk. It is a reputational and safeguarding risk.
Cost blowout. Multi-step reasoning without a routingand budget layer. Predictable and expensive.
Where EdTech companies and universities should start
Narrow scope. One learning loop, one persona, one measurableoutcome. A diagnostic-and-remediation loop for a single subject with a clearsuccess metric is a better first project than a whole-course coach for everylearner.
Human-in-the-loop from day one. The teacher approves theplan. The agent executes. The teacher reviews. Do not skip the review step to launch faster.
Guardrails and grounding before scale, not after. RAG overan approved corpus. Source citation on every generated explanation. Explicit boundaries on what the agent will do for the learner.
Cost architecture from day one. Cost-optimised modelrouting, per-agent budgets, kill switches.
Measure learning outcomes, not just engagement. The trap offirst-wave AI in education was optimising time-on-platform. The second wave hasto prove learning happened. Guardrail metrics matter.
In summary
Agentic AI is not a bigger chatbot. It is a different design pattern for how learning happens with AI in the loop. Assistants respond. Coaches plan, remember, act and adjust. The EdTech companies and universities that pull ahead will be the ones that treat this as an orchestration problem, of agents, humans, data and governance, not as a model choice.
The models are becoming commodity. The loop around them iswhat matters.
How we know this
This article draws on VE3's engagements with EdTech andhigher-education learning platforms, the productised patterns inside PromptX, and public standards including Anthropic's Model Context Protocol, the xAPI and IMS Caliperlearning-events standards, and emerging Responsible AI guidance from UKRI and the OECD AI Principles. Multi-agent framework references include LangChain, LlamaIndex, CrewAI and AutoGen.
About VE3
VE3 is a global technology and enterprise AI consultancy that partners with EdTech companiesand higher-education institutions to build governed AI coaching systems,adaptive learning platforms and multi-agent learning architectures. Our productised platform, PromptX, provides the super-agent layer, cost-optimised model routing and zero-migrationgovernance that makes agentic learning safe to deploy at institutional scale.Our services cover AISolutions, Generative AI Services, Responsible AI Development and Big Data Engineering.
