Metadata Driven AI Pipelines

Long running tasks are always a balance of control and agency. As we slowly move away from the CLI to task managers (Jira, Linear, Beads, etc) we will need to figure out ways to find a good balance between control and agency. If control is pure code and agency is purely agentic, then AI pipelines look to be something in between. Code that executes and validates LLM output, but relies on the intelligence of the model to do a majority of the work. I will show how I am designing the AI pipelines and using metadata from a task manager (Linear in this case) to control which pipeline executes for different types of tasks (bugs, features, etc)