agentic-context-engine

Why we included this project

ACE treats agent mistakes as training data rather than a reason to re-tune weights. Each run is scored by the environment, and a reflector plus a skill manager turn what worked into a readable skillbook of strategies that feeds back into future calls, so the same model gets better at the task without fine-tuning. You can run it as a full pipeline for building agents from scratch, or wrap existing tools like LangChain, browser-use, and Claude Code and learn from their traces. The accumulated skills stay human-readable and model-agnostic, so you can review exactly what changed and carry the strategies across providers if you switch models.