In Progress2025
CognitiveFlow
PythonLangChainQdrantOpenAIFastAPIRedis
Overview
CognitiveFlow is an AI agent routing and orchestration framework built for production reliability. It implements the ReAct pattern with structured tool definitions, vector-backed long-term memory, configurable max-iteration guards, and real-time tracing. Designed for multi-agent workflows where individual agents specialise and hand off tasks through a central router.
Highlights
- Multi-agent routing with specialised sub-agents per domain
- Qdrant vector store for semantic long-term memory across sessions
- OpenTelemetry tracing for full thought/action/observation audit logs
- FastAPI wrapper enabling REST and WebSocket agent execution
Tech Stack
Python 3.12
Core runtime
LangChain
Agent and tool abstractions
Qdrant
Vector store for long-term memory
OpenAI API
LLM backbone (GPT-4o)
FastAPI
REST + WebSocket API layer
Redis
Short-term session memory cache
Challenges
- →Preventing infinite loops in complex multi-step reasoning chains
- →Efficient context window management when conversation history grows
- →Building reliable tool call parsing that handles malformed LLM output gracefully
Outcomes
- ✓Handles 80%+ of benchmark task completions without human intervention
- ✓Average agent run cost under $0.04 on GPT-4o with caching
- ✓Used internally for automated research and summarisation workflows