GEPA

Why it's ranked here

The case for it is adoption, not hype: DSPy, MLflow, and Comet ML's Opik each wired it in as a named optimizer rather than building their own, and the documentation lists production use at companies including Shopify and Databricks. It ships MIT licensed behind one small interface to implement, which lowers the cost of trying it. The central idea, that reading a full execution trace beats collapsing a run to a single score, holds up on its own logic, though the speed and cost comparisons in the README are self-reported and worth treating as marketing until verified independently.

What's good

The adapter model is genuinely thin: connecting a new system means implementing two methods, so the built-in list already covers retrieval pipelines, tool-calling agents, terminal-use agents, and LangChain chains without forcing a rewrite of your own code. Keeping several competing candidates alive at once, rather than settling on a single best-so-far, means a variant tuned for one slice of the task is not thrown away just because it loses on average. Packaging it as an installable agent skill lets a coding agent run the optimization loop on your behalf instead of you scripting it by hand.

Tradeoffs

Because the goal is generality, using it well requires writing your own evaluation function and, for most systems, your own adapter, which is more upfront work than a single-purpose prompt tuner needs. Some optional install extras carry narrow, version-pinned dependency ranges with a documented workaround for a packaging bug in one dependency, though the core package itself has no required dependencies. Every optimization run also calls out to a language model for both the task and the reflection step, so cost scales with how many evaluations the search performs.

How to use it well

This fits teams that already have a way to score an output automatically and are willing to spend a budget of runs searching for a better prompt, code snippet, or agent design instead of hand-tuning one by hand. It pairs naturally with an existing DSPy pipeline, since that integration is the most complete one described. It is a poor fit if you cannot express your goal as an evaluator function, or if you need a single deterministic answer rather than a population of competing candidates to choose from afterward. Budget the reflection calls: each optimization round spends language-model calls on the diagnosis step, not just the task itself.

Observed

License

MIT, per pyproject.toml's license field and the README's license badge.

Packaging

Distributed on PyPI as the gepa package via pip install gepa; the core install declares zero required dependencies in pyproject.toml (dependencies = []).

Python support

Requires Python >=3.10 and <3.15, per pyproject.toml's requires-python.

Interface

Primary interface is a Python library (import gepa, gepa.optimize per README.md); src/gepa/__init__.py exposes optimize, GEPAAdapter, EvaluationBatch, and GEPAResult as the public surface.

Extension mechanism

New systems are connected by implementing a GEPAAdapter with evaluate and make_reflective_dataset methods, described in README.md and exposed from src/gepa/__init__.py.

Built-in adapters

Eight named adapters ship in the repository per README.md: DefaultAdapter, ConfidenceAdapter, DSPy Full Program, Generic RAG, MCP Adapter, TerminalBench, AnyMaths, and LangChain.

Alternate distribution

Also ships as an installable Agent Skill / Claude Code plugin (gepa-optimize-anything) via a plugin marketplace, per README.md, in addition to the pip package.

Test configuration

pyproject.toml declares testpaths = ["tests"] under tool.pytest.ini_options and lists pytest and pyright as test-related dependencies.

Dependency fragility

pyproject.toml's optional extras pin narrow, environment-conditional ranges on litellm, with a comment documenting an incomplete wheel matrix that breaks installs on macOS, Windows, and Python 3.14 unless the version is capped below 1.92.