GEPA vs PromptWizard

GEPA is a free, open-source Python package for optimizing prompts, code, configurations, agent architectures, policies, and other text-based artifacts. Its optimize_anything API connects to any system through an evaluator you write, so DSPy or another framework is not required. GEPA examines execution traces and evaluator feedback, uses language-model reflection to diagnose problems and propose targeted mutations, then selects candidates with Pareto-aware criteria. It can optimize API-only GPT, Claude, and Gemini models without model-weight access. Human-readable traces explain changes and support agent debugging. Built-in adapters cover Default, Confidence, DSPy Full Program, Generic RAG, MCP, TerminalBench, AnyMaths, and LangChain, with integrations including MLflow, Comet ML Opik, Pydantic AI, and cookbook projects. You can limit metric calls and set several stopping conditions. GEPA works through API or self-hosted deployment, installs with pip, and uses the MIT license.

PromptWizard is a free, MIT-licensed open-source framework for automated prompt and example optimization. It generates prompt variations, scores them, critiques their outcomes, and refines instructions over repeated rounds. It can also optimize in-context examples and synthesize examples suited to a task. The described workflows cover prompt optimization without examples, with generated examples, or with training data. Listed datasets include GSM8k, SVAMP, AQUARAT, and Instruction Induction (BBII). Custom datasets use JSONL samples with question and answer fields, and require task-specific answer extraction and evaluation functions, configuration, and data files. PromptWizard can generate chain-of-thought reasoning for in-context examples, with an option to turn that off to reduce prompt length. Installation uses the GitHub repository as a Python package in development mode, with instructions for Windows, macOS, and Linux. Model access setup supports OpenAI API keys and Azure OpenAI endpoints, so access requires credentials for those services. The README reports average optimization times of around 20–30 minutes in experiments on the listed datasets, with timing dependent on the dataset, and notes that human supervision can help tune outputs.