APO — Automatic Prompt Optimizer

Automatically optimize your LLM prompts using gradient-based beam search

APO Paper · AgentLightning Docs · Google ADK Docs

---

A CLI tool that takes your LLM prompt and makes it better — automatically. It uses the **APO (Automatic Prompt Optimization)** algorithm to iteratively critique, edit, and improve prompts through beam search over "textual gradients." Built with [Microsoft AgentLightning](https://github.com/microsoft/agent-lightning) for the optimization loop and [Google ADK](https://github.com/google/adk-python) with [LiteLLM](https://docs.litellm.ai/) for flexible model access across 100+ LLM providers.

How It Works

APO treats prompt optimization like gradient descent, but with natural language instead of numbers:

1. Evaluate — Run your prompt on a batch of tasks and measure performance
2. Critique — An LLM generates "textual gradients" — natural language feedback about what's wrong
3. Edit — Another LLM applies the critique to produce improved prompt candidates
4. Select — Beam search keeps the top-k performing prompts
5. Repeat — Iterate until convergence

*"APO is an iterative prompt optimization algorithm that uses LLM-generated textual gradients to improve prompts through a beam search process."* — [Pryzant et al., 2023](https://arxiv.org/abs/2305.03495)

Dashboard

APO includes a **real-time dashboard** (powered by AgentLightning) that launches automatically at `http://localhost:4747` during optimization.

Tab	What It Shows
**Rol