AI-powered platform to build and orchestrate autonomous AI agent workflows.
Overview
FlowMatch is a visual orchestration engine designed to manage state and execution for autonomous AI agents. It provides a node-based interface for building complex multi-agent workflows, combined with real-time telemetry and execution logging.
The Problem
Orchestrating autonomous AI agents requires maintaining state across distributed nodes while ensuring secure, isolated execution environments. Existing tools were either too rigid (no code) or lacked proper observability for agent hallucinations and failures.
System Architecture
Engineering Trade-offs
Chose eventual consistency for the real-time logging UI (via Redis) to guarantee sub-10ms agent execution latency on the backend, accepting that logs might arrive slightly out of order under extreme load.
Lessons Learned & Future
Production Learnings
Discovered that WebSocket connection drops in serverless environments required building a robust client-side reconnection and event queuing mechanism.
Future Improvements
Migrating the core execution engine to Rust for memory safety and significantly lower baseline resource consumption.
Database Layer
PostgreSQL (Prisma ORM) for persistent workflow definitions and user accounts. Redis for high-throughput, ephemeral agent state and WebSocket pub/sub.
API Design
RESTful architecture for CRUD operations on workflows. WebSocket connections for real-time execution logs and state mutations.
Security Decisions
1. Enforced strict RBAC to prevent cross-tenant workflow execution. 2. Implemented rate limiting on the LLM API proxy to prevent billing exhaustion attacks. 3. Scrubbed sensitive API keys from logs before broadcasting over WebSockets.
Scale & Metrics
Tested to support 500 concurrent agent workflows with <50ms message propagation and 99.9% uptime during load testing.