February 20, 2026
Fork it till you make it
Cord: Coordinating Trees of AI Agents
AI builds its own to‑do list—commenters ask: genius or déjà vu
TLDR: Cord lets an AI build its own task tree, splitting work and tracking dependencies without preset workflows. Commenters argued it’s not new (Claude already does this), predicted frameworks will die as models improve, and debated keeping context versus clean slates—while asking for real performance tests to prove it matters.
Cord promises to let one smart AI break a big job into mini‑tasks on the fly—spawning research, forking analysis, even asking humans for missing info—without the developer micromanaging a fixed workflow. Think “the AI makes its own to‑do tree” instead of you hardcoding every step. But the crowd immediately split. One camp dropped the seen‑it card, pointing to Anthropic’s Claude and its Agent Teams and saying this is old news. Another camp fired off the long‑view hot take: as models get smarter, all these frameworks will vanish and it’ll just be “model + tools.”
Then came the nerd fight over spawn vs fork: does starting a subtask with a clean slate ever beat keeping context? One commenter argued you almost always want context compression, not a blank start. Others dunked on the writing style, calling the post “AI-written” with choppy sentences, and asked for real benchmarks instead of vibes. And yes, there were jokes—someone quipped the agents are named like “frk_ai_8b2e” and live on “news.ycombinator.com,” which is Hacker News itself.
So while Cord’s demo (parallel tasks, dependency tracking) screams “fresh,” the mood is part excitement, part eye‑roll. Everyone wants receipts: performance tests, where it beats Cord clones, and proof this isn’t just fancy task chat.
Key Points
- •The article introduces Cord, a framework where AI agents dynamically construct task trees at runtime.
- •It critiques existing frameworks (LangGraph, CrewAI, AutoGen, OpenAI Swarm, Claude tool-use loops) for requiring predefined coordination and lacking structured parallelism.
- •Cord supports node types (SPAWN, ASK, FORK), explicit dependencies, parallel execution, and context inheritance.
- •An example shows an agent evaluating a REST-to-GraphQL migration by spawning subtasks, querying a human for scale, and forking into comparative analysis.
- •The approach aims to leverage modern models’ planning abilities to discover decomposition instead of hardcoding workflows.