OpenMAIC
OpenMAIC Architecture: From Materials to Durable Multi-Agent Classroom Sessions
Trace OpenMAIC's documented architecture from Next.js workbench and LangGraph orchestration to provider adapters, persistence, artifacts, and exports.

What you will learn
- Map OpenMAIC's observable runtime boundaries
- Evaluate provider and persistence contracts
- Design a traceable architecture review
Before you start
- Basic Git and command-line usage
- Comfort reading a project README
You can explain the project, run its documented first step, and decide what to verify before adopting it.
Key takeaways
- Model OpenMAIC as workbench, agent runtime, artifact tools, and delivery adapters.
- Durable sessions and provider neutrality introduce explicit state and capability contracts.
- Trace materials, state, artifacts, persistence, and exports with reproducible identifiers.
The boundaries visible in the README
OpenMAIC can be reasoned about as four boundaries: a Next.js/React workbench receives a topic and session materials; a LangGraph-backed agent runtime plans and revises a course; course tools produce slides, quizzes, interactive scenes, PBL activities, images, video, or voices; and delivery adapters persist sessions and export artifacts such as .pptx and interactive HTML.
This is an editorial architecture model, not a claim about every internal module. The README names Next.js 16, React 19, TypeScript 5, LangGraph 1.1, and Tailwind CSS 4, plus provider-neutral model, media, search, and storage backends. Pin a commit and inspect the package graph before treating those labels as an implementation contract.
Follow one material through the system
Start with an uploaded document or a web-search result. The session-material layer should normalize metadata and make the source available to the workbench; the agent then turns the learning objective into a plan and invokes course skills. Durable sessions matter here: the v1.0.0 notes say runs can survive restarts, be cancelled, resumed, or steered, so state transitions deserve the same review as UI components.
The artifact boundary is where generated content becomes inspectable. A slide deck, quiz, simulation, or voice asset should carry enough session context to identify its source materials, model/provider route, and generation revision. Export is a separate boundary: a successful chat response does not prove that a .pptx or HTML renderer handled fonts, media, or untrusted input safely.
Provider and persistence adapters
Provider neutrality reduces lock-in but multiplies contracts. Chat, image, TTS, ASR, search, and storage can have different credentials, timeouts, quotas, and data-retention rules. Keep those adapters behind explicit capability checks and record which route produced an artifact; otherwise a fallback can silently change the lesson.
The README highlights a pluggable persistence stack and a one-command Postgres reference deployment. For an architecture review, test a restart between two agent steps, verify that cancellation is idempotent, and compare the restored session with the last emitted artifact. A database connection that works in the happy path is not proof of resumability.
Architecture review checklist
Draw a trace with request ID, material hash, agent state, tool invocation, provider, artifact path, persistence event, and export checksum. Redact document contents and API secrets, but keep identifiers and failure reasons. Then exercise a provider timeout, malformed upload, duplicate resume, and export failure; each should fail at a named boundary and leave a recoverable state.
The most important design choice is keeping generated teaching content reviewable. OpenMAIC's multi-agent classroom can make learning immersive, but subject-matter accuracy, accessibility, and audience suitability remain application-level gates rather than properties guaranteed by orchestration.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Pin a commit and list the runtime, provider, and persistence packages.
- 2
Trace one material from upload/search through an agent step and artifact.
- 3
Restart or cancel mid-run and verify durable state and idempotent resume.
- 4
Export a deck or HTML artifact, record its checksum, and review content and safety.
Copy-ready example
material hash -> session event -> agent state -> tool/provider call
-> artifact metadata -> durable save -> export checksumFrequently asked questions
Does LangGraph alone make a session durable?
No. Durability depends on the application's persistence and recovery implementation; verify it with a restart and a replayable session fixture.
Can every provider produce identical classroom artifacts?
No guarantee. Compare capability, schema, media support, latency, and error behavior per provider and record the selected route.
Sources
- OpenMAIC README (captured 2026-08-31)Source checked 2026-08-31
- OpenMAIC repositorySource checked 2026-08-31