Cord: Coordinating Trees of AI Agents

2026-05-14 · source · source ↗ · raw capture · wiki page

By June Kim. Proposes letting agents dynamically build task trees at runtime rather than requiring developers to predefine coordination structures.

Problem with Existing Frameworks

LangGraph: Static graphs requiring upfront developer decisions
CrewAI: Developer-assigned roles prevent dynamic restructuring
AutoGen: Unstructured conversation lacks dependency tracking
OpenAI Swarm: Linear handoffs without parallelism
Claude's Tool Loops: Single-threaded execution limits context

"Why are we still hardcoding the decomposition?" when models excel at natural planning.

Innovation: Spawn vs. Fork

Two context-flow primitives: - Spawn: Clean slate for independent subtasks; low overhead - Fork: Sibling results injected; expensive but necessary for dependent analysis

Agents make runtime choices about how subtasks inherit context.

Runtime Design

Built on Claude Code CLI with MCP tools backed by SQLite. ~500 lines of Python.

Five primitives: spawn(), fork(), ask(), complete(), read_tree() with dependency resolution.

Behavioral Testing Insights

15 behavioral tests revealed emergent patterns: agents unprompted used read-verify cycles, escalation patterns emerged organically, error recovery through task respawning occurred naturally.

"The passing rate wasn't the insight -- it was that clear tool descriptions plus dependency semantics were sufficient for correct use."

Protocol Independence

The five-primitive protocol is implementation-agnostic. Could extend to multiple LLMs, distributed systems (Postgres), human workers, or direct Claude API.

Key Quote

"Real work isn't one task. It's a tree of tasks with dependencies, parallelism, and context that needs to flow between them."

"The protocol is the contribution. This repo is a proof of concept."