This repository provides an open-source implementation scaffold for Agentic Context Engineering (ACE), a method for self-improving language models through evolving contexts. It targets researchers and engineers seeking to reproduce or build upon ACE, offering a structured framework for context manipulation and model adaptation.

How It Works

ACE structures contexts as playbooks with bullet entries and helpful/harmful counters. Three agentic roles—Generator, Reflector, and Curator—interact via incremental delta updates to refine these contexts. The framework supports both offline and online adaptation loops, enabling multi-epoch training and test-time continual learning for enhanced LLM performance.

Quick Start & Requirements

Requires Python 3.9+ (development used 3.12). The core scaffold has no third-party dependencies. Unit tests can be run via python -m unittest discover -s tests. For practical application, replace DummyLLMClient with production LLM clients (e.g., OpenAI, DeepSeek) and implement a task-specific TaskEnvironment. The script scripts/run_local_adapter.py demonstrates wiring local models (e.g., gpt-oss-20b) into ACE using ace.TransformersLLMClient. Refer to docs/method_outline.md for methodology details.

Highlighted Details

• Core library modules are located in the ace/ directory.
• Implements both offline and online adaptation loops for continuous learning.
• Supports multi-epoch training and test-time continual learning strategies.
• Facilitates easy integration of various LLM backends via LLMClient abstractions.

Maintenance & Community

This project is a reproduction of the ACE method, as the official implementation was not yet released at the time of creation. The author intends to link to the official version upon its availability. No community channels (e.g., Discord, Slack) or specific contributor details are mentioned in the README.

Licensing & Compatibility

The README does not specify a software license. This omission requires clarification for adoption decisions, particularly regarding commercial use or integration into closed-source projects.

Limitations & Caveats

This repository serves as a reproduction and may not perfectly mirror the official ACE implementation. Users should be aware that the official version might differ upon release. The absence of a specified license is a significant caveat for potential adopters.