Self-Refine Agent Strategy

A Dify agent strategy plugin that improves output quality through iterative self-critique and refinement.

What It Does

Self-Refine Agent Strategy implements the Self-Refine (NeurIPS 2023) algorithm, enabling agents to automatically evaluate and improve their own outputs through multiple refinement cycles.

The strategy works by having the LLM act as its own critic:

Execute - Agent generates an initial response
Evaluate - LLM judges the output quality (0-100 score)
Critique - LLM identifies specific issues and improvement opportunities
Refine - Critique is injected into context for the next iteration
Repeat - Process continues until quality threshold is met or max refinements reached

This approach is particularly effective for tasks requiring high-quality outputs such as writing, analysis, code generation, and complex reasoning.

Key Features

Automatic Quality Control - LLM evaluates its own outputs without human intervention
Iterative Improvement - Each refinement cycle builds on previous critiques
Transparent Process - All intermediate steps (execution, evaluation, critique) are visible
Configurable - Control maximum iterations and refinement cycles

Usage

Parameters

Type	Default	Description
model-selector	-	LLM model to use
array[tools]	-	Available tools for the agent
string	-	System instructions for the agent
string	-	User query to process
array[object]	-	Optional context information
number	5	Max iterations per agent execution
number	2	Max refinement cycles (0-5)

Configuration Example

When configuring the agent in Dify:

Select Self-Refine Agent as your strategy
Choose your preferred LLM model
Set to control quality vs. speed tradeoff:
- - No refinement (standard agent behavior)
- - Balanced (recommended for most cases)
- - High quality (best for critical outputs)

Local Development

To test the plugin locally:

Create a file:

Ensure Dify is running locally (via Docker Compose)
Start the plugin:

How It Works

The refinement loop follows this pattern:

The agent automatically stops refining when:

Output quality score reaches 80 or above
Maximum refinement cycles are reached
No significant improvements are detected

Reference

Based on the paper:

Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., ... & Clark, P. (2023).
Self-Refine: Iterative Refinement with Self-Feedback.
NeurIPS 2023. arXiv:2303.17651

License

MIT License