Agentic Context Engine

About Agentic Context Engine

Agentic Context Engine (ACE) is an open-source Python package that adds a persistent learning loop to AI agents. It maintains a Skillbook, a collection of strategies that evolves with every task, and uses three roles to manage it: an Agent that executes tasks with Skillbook strategies, a Reflector that analyzes execution traces to extract what worked and what failed, and a SkillManager that adds, refines and removes strategies. The Recursive Reflector, described as the project's key innovation, writes and executes Python code in a sandboxed environment to search traces for patterns and errors instead of summarizing them in one pass.

ACE installs with uv add ace-framework and offers several runners: ACELiteLLM for a batteries-included agent with ask, learn and save methods; ACE for the full learning loop with batch epochs and evaluation; TraceAnalyser for learning from pre-recorded traces without re-running tasks; and BrowserUse for browser automation. All roles are built on PydanticAI agents with structured output validation and route to more than 100 LLM providers through LiteLLM, with native support listed for OpenAI, Anthropic, Google, Bedrock and Groq. The project states that no fine-tuning, training data or vector database is required. A hosted version is available at kayba.ai.

Key features

• Persistent Skillbook of strategies updated as tasks are executed
• Reflector analyzes execution traces to extract what worked and failed
• Recursive Reflector searches traces by writing Python in a sandboxed environment
• Runners for LiteLLM, core loop, pre-recorded traces and browser automation
• Routes to 100+ LLM providers via LiteLLM; no vector database required
• Installed as a Python package via uv add ace-framework

Agentic Context Engine at a glance