Multi-Agent EDU Stack — a multi-agent educational system for keeping AI enablement current when the method being taught (not just an API surface) obsolesces on a weeks-to-months clock.
Implementation (in development): kaizengrowth/MultiAgentEDUstack. The repo is early-stage scaffolding against the architecture below; expect rapid change. Design rationale and interview context: My Google Interview — Curriculum at Model Speed.
Architecture
Eight stages, source → ship → instrument → decay. Scouts are separate agents (different schedules and judgment), not one agent with five tools.
tier score + story merge] SYN[Synthesis / Digest] TREND[Trend-Forecasting
derivatives not levels] CURR[Curriculum Scaffolding
durable vs frontier] LAB[Lab Generation
state-based validation] EDIT[Editorial Review
pedagogy ∥ technical] SHIP[Shipped Unit] TELE[Behavior Telemetry
Kirkpatrick L3] DECAY[Decay / Deprecation] S1 --> CRED S2 --> CRED S3 --> CRED S4 --> CRED S5 --> CRED CRED --> SYN SYN --> TREND SYN --> CURR TREND --> LAB CURR --> LAB LAB --> EDIT EDIT --> SHIP SHIP --> TELE TELE --> DECAY
Credibility tiers (strict): (1) primary research → (2) lab/vendor primary → (3) named practitioner synthesis → (4) aggregator digest → (5) social chatter. Dedup merges cross-tier duplicates of one story; contradictions escalate to humans, never auto-resolved.
Scaffolding constraints: observable behavioral objectives (backward design); explicit durable-vs-frontier fork; expertise-reversal-aware delivery paths; labs with state-based validation, pinned ephemeral envs, and a CI harness that exercises the lab itself.
Measurement thesis: instrument Kirkpatrick Level 3 in-toolchain (modification rate on accepted suggestions, revert/defect vs staggered cohorts, in-flow micro-assessments). Never individualize for performance review; no composite “fluency index” in year one.
What works well (already validated in the personal stack)
- Source taxonomy + YAML catalog (
data/ai_knowledge_sources.yaml+ filterable explorer) as a living, editable corpus rather than a static bookmark page. - Scholar-alert → digest → atomic note path (shared with the Digital Brain sketch): narrow subfield alerts beat one broad “AI” firehose.
- Transcript-as-corpus for video/podcast: curated channel list → local ASR → same synthesis path as text. Diffable and searchable; watching live does not scale.
- Two-axis review as a design invariant (pedagogy ⊥ technical accuracy): prevents shipping either a correct lab with a bad objective or a beautiful lab teaching stale model behavior.
- Leading-indicator stack for forecasting: arXiv subfield velocity, workshop CFPs, repo star/fork rate, job-posting keyword emergence, citation velocity vs age. Stack ≥2 signals before scaffolding heavyweight content.
In progress / not production-complete
| Stage | Status |
|---|---|
| MultiAgentEDUstack repo | In development — early scaffolding; not production-ready |
| Scout agents (1–5) as scheduled workers | Partial — Scholar/newsletter paths exist in personal tooling; others designed, not shipped as agents |
| Credibility + dedup service | Specced (tier rubric + merge); no durable store/API yet |
| Synthesis / weekly digest | Running in personal research log form; not multi-tenant |
| Trend-forecasting agent | Heuristics documented; no automated derivative scoring job |
| Curriculum scaffolding + lab generation | Design only |
| Editorial review workflow | Process defined; no productized gates |
| Behavior-change telemetry | Measurement model designed; no org-scale instrumentation |
| Decay / deprecation agent | Conceptual; needs unit registry + staleness signals |
This stack is a first-draft generator and discovery system, not a substitute for SME review or learner testing across diverse cohorts.
Evaluate closely / be wary of
- Scout → curriculum automation without human gates. Firehose volume will outpace editorial capacity; default should be propose, not ship.
- Tier-3/4 sources masquerading as ground truth. Aggregators and newsletters often restate lab blogs; keep the “does this teach anything primary sources wouldn’t?” test.
- Lab validation that matches commands instead of state. Teaches incantation, not judgment; breaks under tool churn.
- Forecasting on absolute volume (stars, citations, likes) instead of derivatives — lags the field.
- Telemetry Goodhart risk. If AI-usage metrics enter performance review, the Level-3 signal dies. Cohort/team reporting only.
- Expertise reversal. One-size scaffolding actively harms experts; fork or keep frontier content disposable.
- Unpinned model/tool versions in labs. AI labs rot faster than almost any other lab type; CI the lab steps on a schedule.
- Overclaiming Kirkpatrick Level 4. Attribution across an org is hard; credibility loss is expensive.
Related
- Repo (in development): github.com/kaizengrowth/MultiAgentEDUstack
- Design note: /notepad/ai-curriculum-multi-agent-system/
- Shared ingestion surface: /garden/digital-brain-local-rag/