AI Agents for EdTech: Closing the Gap Chatbots and Copilots Miss

Ship a redesigned assignment builder to a school district in October and you have likely missed the window entirely. EdTech runs on a distribution calendar most other software categories do not share: districts pilot in spring, budget in summer, adopt in August, and a product change that does not land before back to school might not get re-evaluated for another twelve months. Layer onto that a trial funnel where roughly six in ten teachers activate an account during a free pilot but fewer than two in ten convert once the district has to actually pay, and a compliance bar, WCAG 2.1 AA conformance, state ADA procurement requirements, that most B2B SaaS teams never have to clear. Generic AI tools treat "redesign the assignment flow" as a UI task. For an AI agents for edtech need, it is a research problem, an accessibility problem, and a retention problem stacked on top of each other, and getting any one of the three wrong costs a semester, not a sprint.

Why ChatGPT and Cursor miss what edtech actually needs

Ask a generalist chatbot to "improve our assignment creation flow for teachers" and it produces a plausible set of UI recommendations: fewer clicks, a cleaner modal, maybe drag and drop reordering. What it will not do is ask a single teacher how they actually build a differentiated assignment for three reading levels in the twelve minutes they have between passing periods. It has no mechanism for running structured interviews, no way to synthesize a Jobs to Be Done map from what teachers say versus what they actually click, and zero visibility into your product's real trial cohort behavior. The output reads like design advice because it is trained on design advice, not because it understands a classroom, a bell schedule, or a district's procurement cycle.

Cursor and GitHub Copilot solve an adjacent but different problem. They are fast at writing the React component once a human has already decided what the flow should be. An autocomplete tool has no opinion on whether a modal-based assignment builder violates the keyboard navigation requirement a district's accessibility reviewer will check during procurement, and no research function to explain why teachers abandon the platform's built-in differentiation feature in favor of copy-pasting three versions of a worksheet by hand. It will generate a WCAG-non-compliant date picker just as fast as a compliant one, because compliance is not a completion signal to an autocomplete model, it is a domain requirement someone has to hold before the first component gets written.

Draft, Echo, and Proof: the team edtech products actually need

Tonone's Draft is the UX designer on the team: user flows, information architecture, wireframes, interaction design, the work that decides what the assignment builder should be before anyone touches a component library. Draft does not work in isolation on this kind of problem. Ahead of a redesign, Echo, the user researcher, runs structured interviews with the actual users, teachers in this case, and synthesizes what they need using Jobs to Be Done, not satisfaction scores that tell you teachers are "mostly happy" while they quietly build workarounds outside the product. After Draft ships a flow, Proof, the QA engineer, audits it: accessibility conformance, end to end regression, the checks that catch a keyboard trap before a district's accessibility reviewer does. For an edtech team, that research-to-design-to-QA chain is not a nice-to-have process. It is the actual gate between a feature shipping and a feature getting flagged in a district's VPAT review.

Tonone's Echo runs structured teacher interviews and synthesizes Jobs to Be Done before Draft redesigns a single screen, so the workflow fix targets what teachers actually do, not what a survey says they feel.

Echo: finding the workflow mismatch before it becomes a churn reason

The echo-interview skill is built for exactly the situation most edtech teams get wrong: assuming the analytics dashboard tells the whole story. A funnel chart shows that teachers open the assignment builder and abandon it 40% of the time before publishing. It does not show why. Echo runs structured interviews with a representative sample of teachers, using a discussion guide built around actual classroom moments (the Tuesday afternoon prep period, the five minutes before first bell) rather than abstract product questions. The echo-jobs skill then takes the raw interview transcripts and synthesizes them into a Jobs to Be Done map: not "teachers want fewer clicks" but "teachers need to produce three reading-level variants of one assignment in under ten minutes without leaving the tab open during class." That distinction, feature request versus underlying job, is what determines whether Draft redesigns the right thing.

Draft: turning the research into a workflow, not a mockup

With Echo's synthesis in hand, Draft does not start by opening Figma and drawing screens. The draft-flow skill maps the actual task sequence a teacher goes through, prep period to published assignment to graded submission, and identifies where the current product forces context switching that does not match how the job gets done. Only after the flow is settled does the draft-wireframe skill produce the screens, and the draft-review skill checks the resulting design against usability heuristics before it goes anywhere near an engineer. This ordering matters specifically for edtech: a beautiful new assignment modal that still requires six clicks per student reading group solves nothing if the underlying workflow forced that copy-paste behavior in the first place.

Proof: catching the accessibility failure before the district does

Once Draft's redesign is built, proof-audit checks it against WCAG 2.1 AA, the standard nearly every public school district procurement process now requires documentation against, before it reaches a single classroom. This is not a linting pass. It walks the actual keyboard tab order through the new assignment flow, checks screen reader announcements on dynamic content (a common failure point in modal-heavy redesigns), and verifies color contrast on the differentiation tags teachers rely on to tell reading levels apart at a glance. The proof-e2e skill then runs the full assignment creation to grading path end to end, so a regression in the grade sync does not surface for the first time inside a live classroom during a graded assignment.

Tonone's Proof audits WCAG 2.1 AA conformance and runs end-to-end regression on the redesigned flow, catching keyboard traps and grade-sync failures before a district's accessibility reviewer or a live classroom does.

Worked example: Overlook Learning's assignment flow

Overlook Learning runs a K-8 assignment and grading platform live in roughly 400 districts. Product analytics show the free-pilot-to-paid conversion sitting at 19%, well below the 30% the team modeled for this fiscal year, and the churn survey keeps surfacing a vague complaint: "assignment setup takes too long." Nobody on the team can turn that into a fix, because the in-app analytics say the assignment builder's differentiation feature, the one built specifically to save time, is used by only 8% of active teachers. The team hands Overlook's stack to Echo and Draft rather than guessing at another UI tweak.

Echo runs echo-interview with fourteen teachers across three district sizes during a single week, focused on the fifteen minutes before each class period. The pattern that emerges: teachers do not distrust the differentiation feature, they have never found it, because it lives three menus deep behind a "Advanced options" toggle that looks like a settings panel, not a core workflow. Instead, they copy an assignment three times, manually retype grade-level vocabulary into each copy, and upload all three to the gradebook separately, a process that eats 22 minutes of the roughly 30 minutes teachers have for class prep. echo-jobs turns this into a single job statement: "produce three reading-level variants of one assignment, from inside the same screen, in under five minutes, without losing my place in the gradebook."

Draft takes that job statement into draft-flow and redesigns the sequence so leveling happens inline, at the point of assignment creation, instead of behind a settings toggle. The output below is the kind of structured handoff Draft produces before any component gets built, the same discipline the whole cluster runs on:

Overlook Learning, Assignment Flow Redesign
Echo synthesis: teachers copy-paste 3 leveled variants manually, 22 min avg,
differentiation feature undiscovered (buried in Advanced Options, 3 clicks deep).
Job: produce 3 reading-level variants inline, under 5 min, no gradebook context loss.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
draft-flow: new sequence
  1. Teacher opens assignment builder from gradebook (no tab switch)
  2. Base assignment authored once
  3. Leveling control inline on same screen, 3 preset bands
     (below-grade / on-grade / above-grade), auto-adjusts vocabulary
     and reading passage length
  4. Publish targets all 3 bands to the correct student groups
     from the existing gradebook roster, no re-upload step

draft-wireframe: leveling control replaces the buried Advanced
  Options toggle; surfaced as a default step, not an opt-in

draft-review flags: modal-based band selector risks a keyboard
  trap on band switch, escalate to Proof before build
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
proof-audit result: confirmed keyboard trap on band-switch modal,
  screen reader does not announce active band, contrast on "above-
  grade" tag fails AA at 3.8:1. All 3 fixed pre-ship.
proof-e2e result: publish-to-gradebook path passes across all 3
  bands, grade sync verified for each roster group.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Result: assignment setup time drops from 22 min to 6 min in pilot
classrooms. Differentiation feature usage rises from 8% to 61% of
active teachers within the first grading period.

Had Overlook Learning taken this brief to a generalist chatbot, it would have produced a plausible-sounding redesign with no evidence behind it, no accessibility check on the new modal, and no way to know that the actual job was inline leveling, not a faster settings menu. The proof-audit catch alone, a keyboard trap on the redesigned band selector, would have surfaced during the district's own accessibility review months later, at the exact moment it is most expensive to fix: after a signed contract, in front of an IT director who now has a documented compliance failure to report up the chain.

Draft, Echo, and Proof vs the alternatives

None of the three fill the same gap a generalist chatbot or an autocomplete tool fills, and none of them are trying to. The table below breaks down the specific capabilities an edtech product team needs against what each category of tool actually delivers.

Tonone's Draft redesigns the classroom workflow before touching a component, using the job statement Echo synthesized from real teacher interviews, not a satisfaction score.

Install and try

Tonone is free and MIT-licensed. Install it once and Draft, Echo, Proof, and the rest of the roster are available in your Claude Code session. You pay only for the Claude Code token usage during the work.