Inspiration
I wanted to give AI more autonomy without sending every task to the strongest model. Many project tasks are routine, but an AI still needs the right context and a sensible way to hand work to specialized agents. That led to the idea of a local coordinator that understands a project, breaks a goal into steps, and chooses an appropriate route for each one.
What it does
Local Agent Coordinator is designed to use project files as context, create a small plan, and delegate file or document work to registered UGREEN experts. It makes the routing decisions visible and reports sources, gaps, and what actually completed. Complex or uncertain work can be escalated when the runtime supports stronger model tiers; otherwise, the Skill routes by expert capability and does not claim to switch models.
How I built it
I created a Uliya Skill that delegates project-file work to UGREEN’s filesystem expert and supported document extraction to its document expert. I also built a separate, read-only Python prototype that uses a local Ollama model for planning and compares adaptive routing with sending every task to a stronger agent. The submission Skill package contains only Markdown files, as required.
What I learned and the challenges
The hardest part is separating the desired design from capabilities the UGREEN runtime actually exposes. Local model access and model-tier selection still need confirmation, so the Skill includes a capability-based fallback. The official static checker passes, but I have not run the prototype on the Legion or tested the Skill on a UGREEN NAS; its performance and runtime behavior remain to be validated.
