ACE Framework

Agentic Context Engineering — a framework for self-improving language model agents.

ACE enables AI agents to learn from their own execution feedback through three collaborative roles: Agent, Reflector, and SkillManager. Learned strategies accumulate in a Skillbook that makes every subsequent call smarter.

The Learning Loop

graph LR
    S[Sample] --> A[Agent]
    A --> E[Environment]
    E -->|feedback| R[Reflector]
    R -->|analyzes| SM[SkillManager]
    SM -->|updates| SK[Skillbook]
    SK -.->|context| A

Each pass through the loop discovers new strategies, reinforces what works, and prunes what doesn't.

Get Started in 30 Seconds

from ace import ACELiteLLM

agent = ACELiteLLM.from_model("gpt-4o-mini")

answer = agent.ask("If all cats are animals, is Felix (a cat) an animal?")

agent.save("learned.json")

Install

uv add ace-framework

Quick Links

• Getting Started

  ---

  Install the framework and run your first self-improving agent.

  Installation Quick Start

• Concepts

  ---

  Understand the Skillbook, Roles, Insight Levels, and Update Operations.

  Overview

• Guides

  ---

  Build full pipelines, integrate with existing agents, tune prompts.

  Full Pipeline Integration Guide

• Integrations

  ---

  Ready-made runners for LiteLLM, LangChain, browser-use, and Claude Code.

  Integrations Overview

• Pipeline Composition

  ---

  Compose custom pipelines by mixing and matching steps for full control.

  Composing Pipelines Pipeline Engine

Available Runners

Runner	Framework	Use Case
ACELiteLLM	LiteLLM (100+ providers)	Simple self-improving agent
LangChain	LangChain Runnables	Wrap chains/agents with learning
BrowserUse	browser-use	Browser automation with learning
ClaudeCode	Claude Code CLI	Coding tasks with learning
ACE	Full pipeline	Agent + Reflector + SkillManager

Paper

This framework implements the method from:

Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models arXiv:2510.04618