goal-based

Sydney Runkle announced support for rubrics in deepagents, allowing agents to define a clear definition of done and loop until the goal is complete.