Cord: Coordinating Trees of AI Agents
New framework lets AI agents create their own parallel task structures at runtime, not just execute predefined workflows.
Cord is a novel AI agent coordination framework that fundamentally rethinks how multiple agents should work together. Unlike existing solutions like LangGraph (static state machines), CrewAI (fixed roles), AutoGen (unstructured chat), or OpenAI Swarm (linear handoffs), Cord allows agents to dynamically build task trees at runtime. When given a goal like "Should we migrate our API from REST to GraphQL?", a Cord agent analyzes the problem and creates a structured tree of subtasks with proper dependencies, parallelism, and blocking conditions. The system automatically spawns research tasks, creates parallel tracks, blocks tasks until dependencies are met, and even generates questions for human input when needed—all without any pre-defined workflow. This represents a shift from developer-defined coordination to agent-discovered coordination, leveraging modern models' improved planning capabilities.
- Dynamically builds task trees at runtime instead of executing hardcoded workflows
- Enables agents to create parallel subtasks, dependencies, and human interaction points
- Represents a paradigm shift from developer-defined to agent-discovered coordination
Why It Matters
Enables AI agents to handle complex, multi-step projects with adaptive planning rather than rigid, pre-defined workflows.