Instead of letting the LLM write executable code, give it a workflow language you interpret. No sandboxes, no arbitrary code. It already exists: JSONata. #JSONata #AIAgents #Workflow

220,000 tokens and 7 LLM calls down to 5,579 with a JSONata workflow, in a fraction of the time. Interpretable workflows beat code mode, no sandboxes needed. #JSONata #Tokens #SpringAI