Top AI Product

If you’ve spent any time wiring up multi-agent systems, you know the pain. You sketch out a workflow, hardcode agent roles, define handoff patterns, and then pray everything holds together when the task gets messy. Frameworks like LangGraph, CrewAI, and AutoGen all follow this playbook — you, the developer, decide the structure upfront, and agents just execute within those boundaries. [Cord](https://www.june.kim/cord), a new framework by June Kim, flips that on its head.

The core idea is what Kim calls a “coordination tree.” You give it a goal — say, “Should we migrate our API from REST to GraphQL?” — and a root agent reads your prompt, then dynamically decomposes it into a tree of subtasks. No predefined graph, no rigid workflow. The agents figure out dependencies and parallelism on their own. The whole thing is built on just [about 500 lines of Python](https://github.com/kimjune01/cord), using Claude Code CLI, SQLite, and MCP, which is refreshingly lean.

What caught my attention is the two context-flow primitives: **spawn** and **fork**. Spawn gives a child agent a clean slate with only explicit dependencies — like hiring a contractor with a spec. Fork passes along all sibling results — more like briefing a team member who needs the full picture. It’s a simple distinction, but it solves a real problem around context pollution in deep agent hierarchies.

The part that really stands out, though, is how humans fit into this. When an `ask` node triggers, the system pauses and prompts you right in the terminal. Your response gets stored as a result, downstream tasks unblock, and the tree keeps running. You’re not watching from the sidelines — you’re a node in the tree, same as any agent. That felt genuinely different from anything I’ve used before.

Cord hit the [Hacker News front page](https://news.ycombinator.com/item?id=47096466) on February 21st, pulling 152 points and 75 comments. The reception was mixed — some folks questioned whether trees are expressive enough (DAGs came up, naturally), and others pointed out that Claude’s native agent teams already do recursive decomposition. Fair points, but the simplicity of Cord’s approach is exactly what makes it interesting. Not everything needs to be a DAG on day one.

If you’re building anything with multiple agents and you’re tired of hand-drawing workflows, Cord is worth a look. You can get started with `cord run “your goal” –budget 2.0` and see what the tree builds itself into.