An orchestration stack that ran 200+ enterprise marketing AI agents

What it does

KiwiQ compiles Python or JSON graph schemas into LangGraph and executes them via Prefect, offering 24+ reusable node types ranging from LLM calls and conditional routing to web scraping, sandboxed code execution, and human-in-the-loop pauses. A FastAPI frontend wraps the engine with JWT authentication, Stripe billing, and WebSocket streaming, while 27 pre-built workflow definitions cover content creation, diagnostics, and deep research.

The interesting bit

This is not a minimalist library; it is the extracted skeleton of a commercial marketing operation, complete with a billing module and battle scars. By treating workflows as structured data rather than visual diagrams, it forces you to define agent behavior in JSON or Python while the platform handles the noisy infrastructure—RabbitMQ event buses, vector RAG pipelines, and state sharded across PostgreSQL, MongoDB, Weaviate, and Redis.

Key highlights

• 27 production workflow definitions included, from blog generation to LinkedIn scraping and lead scoring
• Multi-provider LLM support: OpenAI, Anthropic, Gemini, Perplexity, Fireworks, and AWS Bedrock
• Real-time human-in-the-loop approvals over WebSocket, plus sandboxed Python code_runner nodes
• Observability through Prefect dashboards, RabbitMQ event streaming, and structured logging
• Ships with a comprehensive CLAUDE.md for Claude Code context, 75+ tests, and 40+ pages of docs

Caveats

• The architecture is intentionally heavy: running it fully requires PostgreSQL, MongoDB, Redis, RabbitMQ, Weaviate, and Prefect
• The 200+ agent claim refers to the original commercial platform; the README does not clarify current open-source adoption metrics
• Workflow examples skew heavily toward marketing automation (blogs, LinkedIn, SEO), so general-purpose users will likely need custom nodes

Verdict

Teams building long-running, multi-step LLM products that need memory, human checkpoints, and commercial auth/billing get a mature head start. Everyone else should probably start with a lighter workflow library.