Skip to content
Nappai Documentation

Critique and Refinement Tool

Critique and Refinement Tool is a reflection tool for an agent. It doesn’t query any service: it makes the agent stop after an important action, such as a search, and write down what it wanted to achieve, what it got, what is missing or wrong, and what its next step will be.

This way the agent spots incomplete or wrong results early and corrects course instead of answering with the first thing it finds. Connect its Tool output to the agent’s Tools input.

When to use it

  • A research agent runs several searches and you want it to check whether each one really answers the question.
  • An agent that queries several sources tends to give incomplete answers.
  • The agent’s tasks have several steps and a mistake in one carries over to the next.
  • You want the run trace to show why the agent made each decision.

Configuration

FieldWhat it does
Tool Name (required)The name the agent sees for the tool. Default: critique_and_refinement_tool.
Tool Description (required)Instructions for the agent: when and how to use the tool. It comes filled in with a text that asks it to use the tool after every important action. You can adapt it to your case.

Output

Tool: the reflection tool. Connect it to the agent’s Tools input. Each time the agent uses it, it gets back its own review organised in four parts: the goal, the result obtained, the critique and the next step.

Example

A consultancy uses an agent that prepares company profiles from internet searches.

  1. Connect a search tool, such as Serp Search API, to the agent’s Tools input.
  2. Add Critique and Refinement Tool and connect it to the Tools input too.
  3. After each search, the agent checks whether it found the data the profile needs. If, for example, the number of employees is missing, it decides to search again with different words.

Good practices

  • Combine it with Planner & Optimizer Tool: one to plan at the start and the other to review each step.
  • Each use of the tool is one more step for the agent: it adds time and model usage. Use it for tasks where quality matters more than speed.