About ReWOO Agent#

The ReWOO (Reasoning WithOut Observation) agent is an advanced agent paradigm that improves efficiency in augmented language models. The ReWOO agent’s implementation follows the ReWOO paper methodology of decoupling reasoning from observations, separating the planning and execution phases, which leads to more efficient tool usage and better token efficiency for reasoning tasks.

The agent uses the NVIDIA NeMo Agent toolkit core library agents and tools to simplify your development experience and deployment. Additionally, you can customize prompts in your YAML config files for your specific needs.

Understanding the ReWOO Agent Phases#

The ReWOO agent implementation is broken down into three phases:

  1. Planning Phase – The agent receives a task and creates a complete plan with all necessary tool calls and evidence placeholders.

    • Plans all steps upfront and uses placeholders (for example, “#E1”, “#E2”), which significantly reduces token consumption.

    • Focuses on logical reasoning, without being distracted by intermediate results generated by the tools in the execution phase.

  2. Execution Phase – The agent executes each step of the plan sequentially, replacing placeholders with actual tool outputs.

    • Replaces placeholders with actual values, eliminating the need to include full tool outputs in each reasoning step.

    • The placeholder system makes the data flow between steps explicit and manageable.

  3. Solution Phase – The agent uses all gathered evidence to generate the final answer.

    • By having a clear plan before execution, the agent is less likely to make incorrect assumptions or get sidetracked by intermediate results before generating the final answer.

To configure your ReWOO agent, refer to Configure the ReWOO Agent.