Orchestrate Autonomous
AI Agent Swarms
Build complex multi-agent system workflows with hybrid vector RAG, Cross-Encoder reranking, Microsoft MarkItDown ingestion, and zero-latency WebSocket telemetry.
Inside AetherFlow
Everything you need to know about our project: what we built, why we engineered it, how it transforms daily workflows, and the mission-critical enterprise problems it solves.
What is AetherFlow?
AetherFlow is an Enterprise Distributed Visual Agentic Workflow & Knowledge Engine. Rather than writing fragile one-off Python scripts or static chaining functions, AetherFlow allows engineering teams to visually compose, execute, monitor, and debug multi-agent AI pipelines on an interactive topological Directed Acyclic Graph (DAG) canvas.
Built with React Flow on the frontend, featuring 7 custom execution nodes: Generative AI Agents (Gemini & OpenAI), Guardrail self-healing validators, full-featured REST API nodes, Hybrid RAG Vector Search, isolated Python 3 / Node.js Code Sandboxes, Condition Routers, and Data Transformers.
Powered by a high-throughput Fastify v5 TypeScript backend, distributed BullMQ workers on Redis, PostgreSQL with Prisma ORM, and zero-latency WebSocket feeds streaming real-time token counts, duration (ms), and step-by-step logs directly onto node handles.
{{node_id.output}} or {{rag_kb.documents}}.Enterprise Background Worker Fleet
Inspect how AetherFlow background workers pull DAG jobs from Redis, execute steps in parallel topological layers, checkpoint intermediate outputs in PostgreSQL, and stream millisecond telemetry over WebSockets.
RagNode: Hybrid Retrieval
Extracts embeddings with BAAI/bge-m3, runs BM25 sparse matching against the vector store, and scores top chunks with AI cross-encoders.
GuardrailNode: DAG Rewind
Catches missing required JSON property 'vat_rate'. Dynamically rewinds the execution graph back to AgentNode with self-correction instructions.
ApiNode: Fail-Fast Sentinel
Missing ERP secret token throws UnrecoverableError. Aborts immediately in 0ms without wasting retry budgets, routing full error logs to DLQ.
Kahn's algorithm evaluates graph dependencies and executes non-conflicting node branches in parallel layers.
Outputs are cached per node. Paused or failed pipelines resume directly without repeating expensive AI prompts.
Missing API keys or schema errors bypass useless retry loops, stopping token waste and freeing threads instantly.
Strict 72-hour TTL auto-eviction prunes dead executions, ensuring Redis memory stays bounded and safe from OOM.
AetherFlow Architecture
Complete full-stack architecture matrix. Click any component to inspect its enterprise specifications and data protocols.
Guardrail Self-Correction Pipeline
Click any node to inspect runtime state, template data, and observe how GuardrailNode rewinds execution on failure.
Deploy Your First Swarm
Connect AetherFlow to your local runtime worker and begin visual workflow execution.