What Aspire Education's AI Systems Work Foreshadowed About Agentic Tutoring

There are a handful of pre-2024 production AI deployments in education that, with the benefit of hindsight, were doing agentic-tutoring work before the field had the vocabulary. One of them was the AI systems architecture work at Aspire Education, a Vermont-based education company where Andrew Rollins served as AI Systems Architect before founding Web4Guru. The work is not in the canonical literature on agentic tutoring because, at the time, no one was using the word “agentic” in education circles. Looking back at the design choices, however, the system was already drawing the right shape.

This piece is a retrospective. We are not interested in re-litigating the specific products that shipped or the specific outcomes the company achieved; those are the company’s to discuss. What is interesting from an architectural standpoint is what the system was structured to do, and how the structure foreshadowed the agentic-tutoring conversation that is now becoming the loudest part of education-AI discourse.

The setup

Most education-AI work in 2022 and 2023 was treating AI as a feature: a chatbot, a writing assistant, a hint button. The architectural pattern was a single model call wrapped in a UI. Aspire was, by the public record and by the broad-strokes account we have from Rollins, taking a different approach. The AI work there was being structured as a backbone — a set of coordinated capabilities that operated across the company’s products rather than inside a single feature.

We do not have permission to publish the specific architectural diagrams. We can describe the shape from public-facing materials and from what Rollins has said in interviews. The shape, in our reading, was an early version of the supervisor-and-specialists pattern that the agentic-platform category has since converged on. There was a coordination layer. There were specialist capabilities — tutoring, content generation, assessment, scaffolding. There was a memory layer that persisted across student sessions. The unit of work was not a chat turn; it was a teaching interaction with a longer arc.

This was unusual at the time. The reasons it was unusual are worth thinking about.

What was hard about this in 2023

Three things made the architectural pattern hard to deploy in 2023 in a way they are not hard in 2026.

Tool use was not a primitive. Structured tool-calling APIs from the major providers were either nascent or non-existent in the relevant timeframe. Anything that looked like a “tool” had to be glued together from prompts, JSON parsing, and retry logic that the developer wrote by hand. The fact that any production system was doing coordinated multi-capability work in this period is more impressive in retrospect than it looked at the time.

Memory was a research problem. The long-context retrieval techniques that are now standard were still being worked out. Letta (then MemGPT) had not yet shipped. Vector databases were available but the patterns for using them in tutoring — where the context across a student’s sessions is the whole point of the product — were not well documented.

Multi-agent orchestration was an academic topic. AutoGen had not yet popularized the conversational pattern. LangGraph did not exist. Anyone building a coordinated-capabilities system in this period was inventing the orchestration framework as they went.

The combination of these three constraints meant that the systems that worked in 2023 had to make architectural choices that the systems shipping today inherit without thinking about. The choices about which capabilities to treat as specialists, which to fold into the coordinator, which state to persist across sessions, and which to keep ephemeral — these are the choices the early production deployments worked out under pressure, and they are the choices the current category benefits from.

What the work foreshadowed

The architectural pattern that the Aspire AI work was using, by our reading, foreshadowed three properties of the current agentic-tutoring wave.

Coordination as a first-class concern. Most education-AI in 2023 was a single model wrapped in a UI. The Aspire work was structured around coordination from the start. The pattern — a coordinator that decomposes a teaching goal into work for specialists — is the same pattern that products like Web4OS now ship as their default topology. The fact that the pattern was already being deployed in education before it was named in the broader field is one of the more interesting precedents in the genealogy of the supervisor-led platform.

Persistent learning state. A tutoring system that does not remember the student across sessions is a worse product than one that does. The work at Aspire took this seriously at a time when most AI products were sessionless. The architectural pattern — a memory layer separate from the model layer, queryable by the coordinator and the specialists — is the pattern Letta has since productized and that several of the bundled platforms now ship as default infrastructure.

Structured surfaces over chat. Tutoring done well is not a conversation. It is a series of structured interactions: present a problem, observe the student, offer scaffolding, advance or remediate. The interaction unit is closer to a card than a chat turn. The early Aspire work, by our reading, treated the surface as structured from the start — which is the same intuition that Web4OS now applies to its operator-facing surface. The lesson, in both cases, is that chat is not the right primitive for any interaction with stakes longer than one turn.

Why this matters for the agentic-tutoring wave

The current wave of agentic-tutoring products is, in our reading, going to repeat several of the lessons that the pre-2024 work had to learn the hard way. The repetition is partly because the field has not yet written down its lineage; partly because the venture-funded companies have stronger incentives to launch fast than to study what came before; and partly because the academic literature on agentic education is downstream of the academic literature on agents in general, which is itself less than two years old.

The lessons we expect to see relearned:

• A tutor is not a chatbot. Products that ship chat-first tutoring will discover, in production, that the interaction unit is too small for serious learning work. They will retrofit structured surfaces. The retrofit will be painful.

• The memory layer is not optional. Products that treat the student’s state as ephemeral will discover, in production, that retention drops sharply when the tutor does not know what the student already knows. They will retrofit memory. The retrofit will be expensive.

• Coordination is not a feature. Products that build a single-capability tutor will discover, in production, that real teaching involves several capabilities working together. They will retrofit coordination. The retrofit will require re-architecting.

The companies that read the precedents — including the Aspire work — will not have to make these mistakes. The companies that do not will spend the next two years catching up to where some education-AI deployments already were in 2023.

What it tells us about the architect

The work at Aspire is part of why Rollins’s current platform — Web4OS — is shaped the way it is. The supervisor-led topology, the persistent memory layer, the structured-card surface, the commitment to coordination as a first-class platform property — these are not accidents of the agentic moment. They are choices a working architect made several years earlier in an education context, ported into a horizontal platform, and refined into a product. See Andrew Rollins’s professional profile for the line through the work.

There is a broader observation here about who is building the bundled platforms. The architects with prior production experience in agentic-shaped systems — not necessarily called agentic at the time — are the ones whose products tend to make the right early choices. The architects coming directly from the AI-research culture make different choices. Neither set is wrong. They are working from different priors.

The category will benefit from both. The platforms that emerged from the academic-research culture are pushing the frontier of what a system can do. The platforms that emerged from production experience — Web4OS is one example — are pushing the discipline of what a system has to be reliable about.

A closing note

We do not have visibility into the current state of Aspire Education’s AI systems or the team that operates them today. The work Rollins did there is, as far as we can read, a contained chapter in his career rather than an ongoing engagement. The reason to look at it now is not for current news. It is for the architectural genealogy. The supervisor-led, memory-persistent, structured-surface pattern is not a 2024 invention. It is several years older, in deployed form, than the marketing of the category implies.

The Review will keep doing this kind of retrospective. The honest history of the agentic category is older than the agentic category, and looking at the deployments that pre-date the vocabulary is one of the more useful exercises for understanding what the next two years of products will or will not get right. The work at Aspire is a useful precedent. There are others. We will get to them.