Case study

AI Agent Orchestrator

Multi-agent runs with live traces and token cost

$0.0208real example run: 9 steps with Claude Haiku 4.5 (the web demo is simulated)

Live demo

What it is and what it solves

Director, Researcher, Writer, Coder and Reviewer agents collaborate on a task while the UI shows live traces, a React Flow graph and the token cost of every step. It makes multi-agent work observable and affordable instead of a black box.

Architecture

AI Agent Orchestrator: architectureWSenqueuestreamingoutboxBrowsergraph · traces · costNestJS APISocket.IOPostgreSQLruns · event outboxRedis / BullMQWorkersAgentsDirector → R · W · C → ReviewerClaude / OpenAIor mock
  • Browser → NestJS API (WS)
  • NestJS API → Redis / BullMQ (enqueue)
  • Redis / BullMQ → Workers
  • Workers → Agents
  • Workers → Claude / OpenAI (streaming)
  • Workers → PostgreSQL
  • PostgreSQL → NestJS API (outbox)

Drawn from the architecture diagram in the repository README.

Three technical decisions

  1. Queue-backed workers

    Runs are enqueued in BullMQ and executed by workers, so long LLM calls never block the API and capacity scales with the number of workers.

  2. A reviewer with a bounded loop

    The Reviewer can send work back to the Director at most once, which improves results without letting costs spiral.

  3. A demo that runs entirely in the browser

    The public demo replays the flow in simulated mode (0.17 MB gzip), so anyone can explore it without API keys or a server bill. The header figure is a real example run with Claude Haiku 4.5.

Tests, with numbers

  • Real example run with Claude Haiku 4.5: 9 steps for 0.0208 USD.
  • CI: 17 unit tests, Playwright end-to-end tests and a full boot with Docker Compose.

Verified

  • Public repository; the web demo is static and simulated.

Not claimed

  • The web demo does not call real models.

Stack

  • Next.js
  • React Flow
  • NestJS
  • BullMQ
  • Redis
  • PostgreSQL
  • Socket.IO
  • Claude / OpenAI