Kembali ke Studio

Arsitektur Multi-Agent

Sistem modular, event-driven, dan queue-based: 19 AI Agent stateless yang berkolaborasi hanya melalui Project Memory, dikendalikan Workflow Orchestrator dengan retry otomatis, resumable workflow, logging lengkap, dan monitoring real-time.

AI Agent
19
Wave eksekusi
13
Estimasi pipeline
250s
Durasi output
20 detik

Diagram alur eksekusi

Topological waves — agent dalam satu wave berjalan paralel

Wave 1 · 1 agent · 2s

Project Manager AI

Wave 2 · 1 agent · 8s

Script Analyst AI∥

Wave 3 · 1 agent · 7s

Story Optimizer AI

Wave 4 · 2 agent · 15s

Storyboard AIVoice Director AI∥

Wave 5 · 5 agent · 9s

Character Director AI∥Environment Director AI∥Music Director AI∥Sound Effect AI∥Subtitle AI∥

Wave 6 · 1 agent · 14s

Prompt Engineer AI

Wave 7 · 1 agent · 40s

Image Director AI∥

Wave 8 · 1 agent · 10s

Motion Director AI∥

Wave 9 · 1 agent · 60s

Video Director AI∥

Wave 10 · 1 agent · 25s

Video Composer AI

Wave 11 · 1 agent · 12s

Quality Control AI

Wave 12 · 2 agent · 18s

Thumbnail Designer AI∥SEO Specialist AI∥

Wave 13 · 1 agent · 30s

Export Manager AI

Daftar AI Agent & kontrak data

Setiap agent membaca dan menulis hanya melalui Project Memory

Struktur Project Memory

Shared state terversi: artefak, status, revisi, error, log

ProjectMemory {
  projectId, version, createdAt, updatedAt
  artifacts: {
    metadata, narration, analysis, optimizedScript, storyboard,
    characterBible, environmentLibrary,
    imagePrompts, videoPrompts, thumbnailPrompts,
    images, motion, videos,
    voice, music, sfx, subtitles,
    timeline, qc, thumbnails, seo, exports
  }
  tasks:     { [agentId]: { status, attempts, durationMs, error } }
  revisions: [{ at, agentId, key, note }]
  errors:    [{ at, agentId, attempt, message, fatal }]
  logs:      [{ at, level, agentId, message }]
}

Setiap tulisan menaikkan version dan mencatat revisi, sehingga workflow dapat dilanjutkan dari langkah terakhir walau sesi terputus.

Workflow Orchestrator

Queue-based scheduler dengan dependensi, paralelisasi, dan resume

  1. 1.Menyusun graf dependensi agent menjadi wave topologis.
  2. 2.Menjalankan agent paralel dalam satu wave, sekuensial bila saling bergantung.
  3. 3.Memvalidasi input dari Project Memory sebelum agent dieksekusi.
  4. 4.Retry otomatis dengan exponential backoff sampai batas per agent.
  5. 5.Menghentikan pipeline saat error fatal dan mencatat penyebabnya.
  6. 6.Resume: hanya menjalankan agent yang belum berstatus done.
  7. 7.Rework loop: QC mengembalikan pekerjaan ke agent terkait saja.
  8. 8.Emit event real-time untuk monitoring, ETA, dan log dashboard.

Monitoring real-time

Simulasi orkestrasi penuh — retry, resume, dan log tercatat

Project SSA-MS6NA925 · v1 · sisa estimasi 300s

0%

Project Manager AIqueued
Script Analyst AIqueued
Story Optimizer AIqueued
Storyboard AIqueued
Character Director AIqueued
Environment Director AIqueued
Prompt Engineer AIqueued
Image Director AIqueued
Motion Director AIqueued
Video Director AIqueued
Voice Director AIqueued
Music Director AIqueued
Sound Effect AIqueued
Subtitle AIqueued
Video Composer AIqueued
Quality Control AIqueued
Thumbnail Designer AIqueued
SEO Specialist AIqueued
Export Manager AIqueued

Error handling & recovery

Gagal satu agent tidak pernah mengulang seluruh pipeline

Skalabilitas & rekomendasi implementasi

Tambah agent baru tanpa menyentuh core system

Agent masa depan (plug-in)

  • AI Translator — setelah subtitle → menulis subtitles.locales
  • AI Lip Sync — setelah video-director → menulis videos.lipsync
  • AI Voice Cloning — menggantikan voice-director via provider adapter
  • AI Scene Expansion — setelah storyboard → memperluas jumlah scene
  • AI B-Roll Generator — paralel dengan image-director → menulis images.broll
  • AI Trend Analyzer — sebelum script-analyst → memperkaya analysis
  • AI Analytics Advisor — setelah export-manager → menulis seo.insights

Stack produksi

  • Frontend: React 19 + TanStack Start + Tailwind v4 (aktif).
  • Agent runtime: server function per agent, stateless & idempoten.
  • Project Memory: Postgres (JSONB artifacts) + object storage untuk media.
  • Queue & durable execution: worker antrean event-driven dengan retry terkelola.
  • AI Gateway: satu adapter model untuk teks, gambar, dan suara.
  • Observability: log terstruktur, metrik durasi per agent, alert error fatal.