Cord: Coordinating Trees of AI Agents
4 days ago
- #Multi-Agent Systems
- #Dynamic Task Management
- #AI Coordination
- AI agents excel at focused tasks but struggle with complex, multi-task workflows.
- Existing multi-agent frameworks (LangGraph, CrewAI, AutoGen, OpenAI Swarm) require predefined coordination structures, limiting flexibility.
- Cord introduces a dynamic approach where agents build task trees at runtime, deciding dependencies and parallelism autonomously.
- Key innovation: 'spawn' vs 'fork' for context-flow—spawn starts fresh, fork inherits context from siblings.
- Cord uses Claude Code CLI with MCP tools and SQLite, enabling agents to decompose tasks, ask human questions, and manage dependencies.
- The protocol is model-agnostic, supporting multiple LLMs, databases, and even human workers.
- Cord's effectiveness was validated through tests showing Claude's intuitive understanding of coordination without prior training.
- Available as a proof of concept on GitHub, requiring Claude Code CLI for execution.