ReWOO Orchestration: Static Plan Validation Before Agent Tool Execution

AWS Machine Learning Blog·high signal

Separates agent execution into three discrete auditable stages: a Planner emits a declarative program (naming tools and arguments in a structured format like `#E1 = ToolName[key=value]` without executing them), a Worker runs only authorized tools maintaining an evidence ledger, and a Solver synthesizes answers from evidence only without calling tools. This static-validation-before-execution model enables policy enforcement on the full plan before any state mutation occurs. Available in the AWS Strands Agents SDK with built-in plan serialization; prevents the runaway tool-call problem that plagues ReAct-style loops where the LLM decides what to call at each step.