A new tool called Cord enables AI agents to dynamically build task trees without predefined workflows. Instead of static graphs, agents like Claude autonomously split complex goals into subtasks, parallelize work, and inject human context when needed. In a test, an agent evaluating a REST-to-GraphQL migration audit generated a 47-endpoint analysis, researched trade-offs, and blocked analysis on human input about user scale. This contrasts with frameworks like LangGraph (static graphs) or CrewAI (role-based rigidity), which force developers to hardcode dependencies. Cord’s innovation lies in its runtime decision-making: agents choose between `spawn` (clean-slate subtasks) and `fork` (context-rich tasks) based on dependencies.

For example, it spawned parallel research threads but forked analysis to inherit prior findings. The system uses Claude Code CLI and SQLite for real-time coordination, with agents unaware of their place in the tree. When tested, Claude autonomously used `read_tree()` to map progress and structured subtasks without explicit instruction. **The result?

AI that mirrors human problem-solving — decomposing projects into 3-6 subtasks, asking precise questions, and escalating roadblocks. This isn’t just theoretical: Cord’s protocol** (dependency resolution, authority scoping) could power everything from software engineering to enterprise workflows. By letting agents build their own coordination trees, it unlocks adaptive AI systems that evolve with task complexity.