# GEPA Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/gepa
> Markdown URL: https://aitinkerers.org/technologies/gepa.md
> Technology record last updated: 2026-02-26T12:50:37Z
> Generated: 2026-09-22T00:37:47Z

GEPA (Genetic-Pareto) is a prompt optimization framework for LLMs: it uses natural language reflection to evolve instructions, achieving up to 35x greater sample efficiency than traditional Reinforcement Learning (RL) methods.

This is a paradigm shift in Large Language Model (LLM) tuning. GEPA (Genetic-Pareto) employs an evolutionary search loop: it uses a powerful reflection model to analyze execution traces, diagnose problems, and propose targeted prompt updates. This method leverages rich, textual feedback—not sparse scalar rewards—to learn high-level rules from trial and error. On complex reasoning tasks, GEPA consistently outperforms baselines like Group Relative Policy Optimization (GRPO), requiring up to 35 times fewer rollouts to achieve superior performance. The framework is available via a simple `pip install gepa` command, with direct integration into systems like DSPy.

- Official technology site: https://github.com/gepa-ai/gepa
- Public AI Tinkerers demos and talks: 3
- Result page: 1 of 1

## Recent Public Talks and Demos

### [Agents Building Agents: Reflective Optimization Loops](https://toronto.aitinkerers.org/talks/rsvp_4gS7qartFFk)

I'll show how an AI agent can build and optimize another AI agent, using reflective optimization to find issues, optimize evals, iterate on architecture, find the optimal prompt/model, and more. We've built a system with multiple levels of reflective optimization for agent development. - GEPA: reflective prompt optimization - Synthetic eval generation: going from a 1-off bug to an proper eval you can use in reflective optimization - Expanding reflective optimization beyond prompts: model selection, tool use, subagents -- reflective optimization can drive all levels of agent optimization.

- Event context: AI Tinkerers Toronto - April 2026 - hosted by Shopify — 2026-04-29 — Toronto
- Public talk page: https://toronto.aitinkerers.org/talks/rsvp_4gS7qartFFk

### [Encoding Engineering Judgment into AI Coding Agents](https://toronto.aitinkerers.org/talks/rsvp_d8pBUS_N4jA)

LLMs are great at writing code, but they mostly follow patterns they learned from random internet repos. That’s fun until those patterns start leaking into your production codebase. In this talk I’ll demo Mystro, a system that lets you encode your own coding practices and use them to guide an AI coding agent. Instead of hoping the model does the right thing, you can test, evaluate, and retry automatically until it starts following your rules. It's using the same underlying algo (GEPA) as autoresearch from Karparthy that was recently released. Think less vibes, more engineering policy. A quick look at what it might take to actually run coding agents inside real codebases.

- Event context: AI Tinkerers Toronto - March - hosted by Mozilla! — 2026-03-25 — Toronto
- Public talk page: https://toronto.aitinkerers.org/talks/rsvp_d8pBUS_N4jA

### [Automatically Optimizing Agents for Deep Research](https://amsterdam.aitinkerers.org/talks/rsvp_HV9939ok-vg)

Research we did comparing different methodologies for self-optimizing complex agent systems. In particular we evaluated a Deep Research architecture and showed that these auto-optimized agents can outperform human-created ones.

- Event context: AI Tinkerers Amsterdam — October Edition: Agents in production — 2025-10-10 — Amsterdam
- Public talk page: https://amsterdam.aitinkerers.org/talks/rsvp_HV9939ok-vg

## Related Technologies

- [Agent Skills](https://aitinkerers.org/technologies/agent-skills) ([Markdown](https://aitinkerers.org/technologies/agent-skills.md)) — 4 public demos
- [AI coding agents](https://aitinkerers.org/technologies/ai-coding-agents) ([Markdown](https://aitinkerers.org/technologies/ai-coding-agents.md)) — 3 public demos
- [Autoresearch](https://aitinkerers.org/technologies/autoresearch) ([Markdown](https://aitinkerers.org/technologies/autoresearch.md)) — 1 public demo
- [Codex](https://aitinkerers.org/technologies/codex) ([Markdown](https://aitinkerers.org/technologies/codex.md)) — 44 public demos
- [GLM](https://aitinkerers.org/technologies/glm) ([Markdown](https://aitinkerers.org/technologies/glm.md)) — 3 public demos
- [Haiku](https://aitinkerers.org/technologies/haiku) ([Markdown](https://aitinkerers.org/technologies/haiku.md)) — 7 public demos
- [Kiln AI](https://aitinkerers.org/technologies/kiln-ai) ([Markdown](https://aitinkerers.org/technologies/kiln-ai.md)) — 1 public demo
- [LanceDB](https://aitinkerers.org/technologies/lancedb) ([Markdown](https://aitinkerers.org/technologies/lancedb.md)) — 5 public demos
- [LiteLLM](https://aitinkerers.org/technologies/litellm) ([Markdown](https://aitinkerers.org/technologies/litellm.md)) — 21 public demos
- [LLM-as-a-judge](https://aitinkerers.org/technologies/llm-as-a-judge) ([Markdown](https://aitinkerers.org/technologies/llm-as-a-judge.md)) — 1 public demo
- [OpenRouter](https://aitinkerers.org/technologies/openrouter) ([Markdown](https://aitinkerers.org/technologies/openrouter.md)) — 34 public demos
- [Opus](https://aitinkerers.org/technologies/opus) ([Markdown](https://aitinkerers.org/technologies/opus.md)) — 7 public demos
- [Python](https://aitinkerers.org/technologies/python) ([Markdown](https://aitinkerers.org/technologies/python.md)) — 662 public demos
- [RAG](https://aitinkerers.org/technologies/rag) ([Markdown](https://aitinkerers.org/technologies/rag.md)) — 147 public demos
- [TextGrad](https://aitinkerers.org/technologies/textgrad) ([Markdown](https://aitinkerers.org/technologies/textgrad.md)) — 1 public demo