Agentic Context Engine

Agentic Context Engine, usually shortened to ACE, is an open source Python framework from Kayba that lets AI agents improve from their own experience without fine tuning, training data or a vector database. It keeps a Skillbook, a persistent collection of strategies that evolves with every task, and feeds those strategies back into the agent on later runs. The framework is aimed at developers who run agents in production and want failures to turn into lasting fixes. Three roles manage the learning loop: the Agent executes tasks, the Reflector analyzes execution traces to find what worked and what failed, and the SkillManager adds, refines and removes strategies. The Recursive Reflector writes and runs Python in a sandbox to search traces for patterns rather than summarizing them in one pass. Runners cover LiteLLM models, pre recorded traces, browser automation, LangChain chains and Claude Code tasks, and an MCP server brings it into the IDE. A composable pipeline engine lets teams build custom learning sequences, and existing agent logs can be mined for strategies directly. The framework itself is free to install from the public repository. Kayba also sells a hosted product that connects to Sentry, PostHog or OpenTelemetry, investigates each agent error, proposes a scored fix as a pull request and Slack message, and tracks how the fix performs against evals built from real failures. Kayba does not publish prices for that service; access starts with a demo booking. The listed website now leads with the commercial Kayba product rather than the open source framework, and the method is based on the ACE research paper from Stanford and SambaNova.