What is GEPA in the context of DSPy?

GEPA (Generalized Evolution Prompting for Agents) is a DSPy optimizer that improves agent performance by reflecting on execution traces and textual feedback to evolve better instructions.

What kind of metrics are required for GEPA?

GEPA requires a metric that returns a tuple containing both a numerical score and a textual feedback string explaining the quality of the result.

How does this skill differ from other DSPy optimizers like MIPRO?

While MIPRO focuses on prompt search, GEPA is specifically designed for agentic systems, utilizing execution trajectories and feedback metrics to understand why an agent failed before suggesting improvements.

Do I need a specific model for the reflection phase?

Yes, GEPA works most effectively when using a strong reasoning model (such as GPT-4 or Claude 3.5 Sonnet) as the reflection_lm to analyze the agent's behavior.

Can this skill optimize the tools my agent uses?

Yes, by enabling tool optimization, GEPA can refine the docstrings and descriptions of your tools to help the agent understand when and how to use them correctly.

DSPy GEPA Reflective Optimizer

Name: DSPy GEPA Reflective Optimizer
Author: OmidZamani

byOmidZamani

•

Ciencia de Datos y ML

Optimizes complex AI agents and ReAct systems using LLM-driven reflection on execution trajectories and textual feedback.

The dspy-gepa-reflective skill implements the Generalized Evolution Prompting for Agents (GEPA) optimizer within the DSPy framework to enhance multi-step agentic systems. Unlike traditional optimizers that rely on simple numerical scores, this skill utilizes rich textual feedback and full execution traces to identify failure points. By leveraging a high-reasoning 'reflection LM,' it performs a Pareto-based evolutionary search to refine both agent instructions and tool descriptions. It is an essential tool for developers building sophisticated agents where reasoning steps, tool usage, and instruction clarity are critical to performance.

Características Principales

01Deep integration with high-reasoning models for agent failure analysis

02Textual reflection on execution traces for nuanced agent optimization

0331 GitHub stars

04Pareto-based evolutionary search for superior instruction candidates

05Joint optimization of predictor instructions and tool descriptions

06Advanced support for complex multi-step ReAct agent workflows

Casos de Uso

01Optimizing tool docstrings to ensure agents select the correct functions for specific tasks

02Refining the reasoning steps and tool calls of a complex autonomous research agent

03Improving agent reliability by analyzing execution logs to generate improved prompts

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add omidzamani/dspy-skills dspy-gepa-reflective

For use in Claude.ai and ChatGPT

Download Skill