AI agents that get smarter with every task
â Star this repo if you find it useful!
ACE enables AI agents to learn from their execution feedbackâwhat works, what doesn'tâand continuously improve. No fine-tuning, no training data, just automatic in-context learning.
The framework maintains a Skillbook: a living document of strategies that evolves with each task. When your agent succeeds, ACE extracts patterns. When it fails, ACE learns what to avoid. All learning happens transparently in context.
- Self-Improving: Agents autonomously get smarter with each task
- 20-35% Better Performance: Proven improvements on complex tasks
- 49% Token Reduction: Demonstrated in browser automation benchmarks
- No Context Collapse: Preserves valuable knowledge over time
- Direct your favorite coding agent (Cursor, Claude Code, Codex, etc) to Quick Start Guide
- Prompt away!
pip install ace-frameworkexport OPENAI_API_KEY="your-api-key"from ace import ACELiteLLM
agent = ACELiteLLM(model="gpt-4o-mini")
answer = agent.ask("What does Kayba's ACE framework do?")
print(answer) # "ACE allows AI agents to remember and learn from experience!"Done! Your agent learns automatically from each interaction.
â Quick Start Guide | â Setup Guide
Wrap your existing agent (browser-use, LangChain, custom) with ACE learning. Your agent executes tasks normally while ACE analyzes results and builds a skillbook of effective strategies.
The Skillbook acts as an evolving system prompt that automatically improves based on execution feedbackâno manual prompt engineering required.
Run coding tasks with Claude Code while ACE learns patterns from each execution, building expertise over time for your specific codebase and workflows.
Create new agents with built-in learning for customer support, data extraction, code generation, research, content creation, and task automation.
A challenge where LLMs often hallucinate that a seahorse emoji exists (it doesn't).
In this example:
- The agent incorrectly outputs a horse emoji
- ACE reflects on the mistake without external feedback
- On the second attempt, the agent correctly realizes there is no seahorse emoji
Online Shopping Demo: ACE vs baseline agent shopping for 5 grocery items.
In this example:
- ACE learns to navigate the website over 10 attempts
- Performance stabilizes and step count decreases by 29.8%
- Token costs reduce 49.0% for base agent and 42.6% including ACE overhead
â Try it yourself & see all demos
In this example, Claude Code is enhanced with ACE and self-reflects after each execution while translating the ACE library from Python to TypeScript.
Python â TypeScript Translation:
| Metric | Result |
|---|---|
| Duration | ~4 hours |
| Commits | 119 |
| Lines written | ~14k |
| Outcome | Zero build errors, all tests passing |
| API cost | ~$1.5 (Sonnet for learning) |
ACE integrates with popular agent frameworks:
| Integration | ACE Class | Use Case |
|---|---|---|
| LiteLLM | ACELiteLLM |
Simple self-improving agent |
| LangChain | ACELangChain |
Wrap LangChain chains/agents |
| browser-use | ACEAgent |
Browser automation |
| Claude Code | ACEClaudeCode |
Claude Code CLI |
| Opik | OpikIntegration |
Production monitoring and cost tracking |
â Integration Guide | â Examples
Based on the ACE research framework from Stanford & SambaNova.
ACE uses three specialized roles that work together:
- Agent - Creates a plan using learned skills and executes the task
- Reflector - Analyzes what worked and what didn't after execution
- SkillManager - Updates the skillbook with new strategies based on reflection
Important: The three ACE roles are different specialized prompts using the same language model, not separate models.
ACE teaches your agent and internalizes:
- Successes â Extract patterns that work
- Failures â Learn what to avoid
- Tool usage â Discover which tools work best for which tasks
- Edge cases â Remember rare scenarios and how to handle them
The magic happens in the Skillbookâa living document of skills that evolves with experience. Key innovation: All learning happens in context through incremental updatesâno fine-tuning, no training data, and complete transparency into what your agent learned.
---
config:
look: neo
theme: neutral
---
flowchart LR
Skillbook[("`**Skillbook**<br>(Evolving Context)<br><br>â˘Strategy Skills<br> Helpful skills <br> Harmful patterns <br> Neutral observations`")]
Start(["**Query** <br>User prompt or question"]) --> Agent["**Agent** <br>Executes task using skillbook"]
Agent --> Reflector
Skillbook -. Provides Context .-> Agent
Environment["**Task Environment**<br>Evaluates answer<br>Provides feedback"] -- Feedback+ <br>Optional Ground Truth --> Reflector
Reflector["**Reflector**<br>Analyzes and provides feedback what was helpful/harmful"]
Reflector --> SkillManager["**SkillManager**<br>Produces improvement updates"]
SkillManager --> UpdateOps["**Merger** <br>Updates the skillbook with updates"]
UpdateOps -- Incremental<br>Updates --> Skillbook
Agent <--> Environment
- Quick Start Guide - Get running in 5 minutes
- Setup Guide - Installation, configuration, providers
- Integration Guide - Add ACE to existing agents
- API Reference - Complete API documentation
- Complete Guide to ACE - Deep dive into concepts
- Prompt Engineering - Advanced prompt techniques
- Agentic System Prompting - Automatically generate prompt improvements from past traces
- Examples - Ready-to-run code examples
- Benchmarks - Evaluate ACE performance
- Changelog - Recent changes
We love contributions! Check out our Contributing Guide to get started.
Based on the ACE paper and inspired by Dynamic Cheatsheet.
If you use ACE in your research, please cite:
@article{zhang2024ace,title={Agentic Context Engineering},author={Zhang et al.},journal={arXiv:2510.04618},year={2024}}â Star this repo if you find it useful!
Built with â¤ď¸ by Kayba and the open-source community.